social investments, informal risk sharing, and …

SOCIAL INVESTMENTS, INFORMAL RISK SHARING, AND

INEQUALITY

ATTILA AMBRUS†, ARUN G. CHANDRASEKHAR‡, AND MATT ELLIOTT?

Abstract. This paper studies the formation of risk-sharing networks through costly social

investments, identifying endogenous inequality and a misallocation of resources as possible

downsides to such agreements. First, individuals invest in relationships to form a network.

Next, neighboring agents negotiate risk-sharing arrangements, in a generalized version of

the model in Stole and Zwiebel (1996). This results in the social surplus being allocated

according to the Myerson value. In particular, more centrally connected individuals receive

higher shares. We find a novel trade-off between efficiency and equality. The most stable ef-

ficient network, which minimizes incentives to overinvest, also generates the most inequality.

When individuals are split into groups and relationships across groups are more costly but

incomes across groups are less correlated, there is never underinvestment into social connec-

tions within group, but underinvestment across groups is possible. More central agents have

better incentives to form across-group links, reaffirming the efficiency inequality trade-off.

Evidence from 75 Indian village networks is congruent with our model.

Date: July 13, 2015.We thank the National Science Foundation, SES-1155302, for funding the research that collected the data weuse. Ambrus, Chandrasekhar and Elliott acknowledge financial support from National Science Foundation(NSF) grant “The economic benefits of investing into social relationships,” SES-1429959. We thank NageebAli, Ben Golub, Matt Jackson, Willemien Kets, Cynthia Kinnan, Rachel Kranton, Peter Landry, HoracioLarreguy, Melanie Morten, Kaivan Munshi, Heinrich Nax and Luigi Pistaferri for helpful comments. Wealso thank seminar participants at the Calvo-Armengol prize conference, the Harvard-MIT theory seminar,Oxford, Cambridge, Warwick, Microsoft Research, 2012 ThRed conference in Budapest, Syracuse, Ohio StateUniversity, Duke, the Tinbergen Institute and the Santa Fe Institute.†Department of Economics, Duke University; email: [email protected].‡Department of Economics, Stanford University, and NBER; email: [email protected].?Division of Humanities and Social Sciences, California Institute of Technology; email: [email protected].

SOCIAL INVESTMENTS, INFORMAL RISK SHARING, AND INEQUALITY 1

1. Introduction

In the context of missing formal insurance markets and limited access to lending and bor-

rowing, incomes may be smoothed through informal risk-sharing agreements that utilize social

connections. A large theoretical and empirical literature studies how well informal arrange-

ments replace the missing markets.1 However, the existing literature does not investigate

a potential downside to these agreements: if people’s network position affects the share of

surplus generated by risk sharing they appropriate, social investments may be distorted and

inequality may endogenously arise. These questions can only be addressed in a framework

with endogenous network formation.2

In this paper we provide an examination of these issues, by considering a simple two stage

model. In the first stage villagers invest in costly bilateral relationships, knowing that in the

second stage they will reach informal risk-sharing agreements. These agreements determine

how the surplus generated by risk sharing is distributed, and they depend on the endogenous

structure of the social network from the first stage. In this way we elucidate new costs as-

sociated with informal risk-sharing. Once incomes have been realized, risk sharing typically

reduces inequality by smoothing incomes. Nevertheless, asymmetric equilibrium networks

generate inequality in expected utilities terms. Agents occupying more advantageous posi-

tions in the social network appropriate considerably more of the benefits generated by risk

sharing. Indeed, seeking to occupy such positions in the network might lead villagers to

spend too much time building social capital. Alternatively, if risk sharing with one neighbor

generates positive spillovers for other neighbors, there can be too little investment in forming

relationships undermining the effectiveness of informal risk-sharing.

Empirical work suggests that both underinvestment and overinvestment in social capital

are possible, in different contexts. Austen-Smith and Fryer (2005) cites numerous references

from sociology and anthropology, suggesting that members of poor communities allocate

inefficiently large amounts of time to activities maintaining social ties, instead of productive

activities. In contrast, Feigenberg et al. (2013) find evidence in a microfinance setting that it

is relatively easy to experimentally intervene and create social ties among people that yield

substantial benefits. One explanation for this finding is that there is underinvestment in

social relationships.

It is important to study the above aspects of informal risk sharing, both to put related aca-

demic work (which often takes social connections to be exogenously given) into context and

to guide policy choices. Consider, for example, microfinance. Two central aims of such inter-

ventions are to increase the efficiency of investment decisions by providing better access to

1An incomplete list of papers includes Rosenzweig (1988), Fafchamps (1992), Coate and Ravallion (1993),Townsend (1994), Udry (1994), Ligon, Thomas and Worrall (2002), Fafchamps and Gubert (2007), Bloch,Genicot, and Ray (2008), Angelucci and di Giorgi (2009), Jackson, Rodriguez-Barraquer and Tan (2012),Ambrus, Mobius and Szeidl (2014).2Previous works that do consider the network formation problem include Bramoulle and Kranton (2007a,b)in the theoretical literature and Attanasio et al. (2012) in the experimental literature. For a related paperoutside the networks framework, see Glaeser et al. (2002).


capital and to reduce inequality. Clearly the value of microfinance then depends on whether

informal risk sharing promotes equality or inequality and whether there is underinvestment,

overinvestment or efficient investment in social connections. If there is overinvestment, mi-

crofinance has a greater scope for efficiency savings in terms of reducing people’s allocation of

time into social investments. With underinvestment, however, it has more scope for smooth-

ing incomes. If there is neither under- nor overinvestment, it tells us that informal risk sharing

is working relatively well as a second-best solution. Understanding which regime applies can

help anticipate policy implications and evaluate welfare impacts of interventions.

To explore efficiency and inequality and capture the relevant trade-offs, we impose several

simplifying assumptions that keep the analysis tractable: agents have CARA utilities, their

income realizations are jointly normal, and that there are no contracting problems. In the

first stage, agents choose with whom to form connections. Link formation is costly, as in

Myerson (1991) and Jackson and Wolinsky (1996). In the second stage, pairs of agents who

have formed a connection commit to a bilateral risk-sharing agreement (transfers contingent

on income realizations). We assume that these agreements can be perfectly enforced.

We investigate agreements satisfying two simple properties. First we require agreements to

be pairwise efficient, in that no pair of directly connected agents leave gains from trade on the

table.3 Second, following Stole and Zwiebel (1996), we require the agreements to be robust to

“split-the-difference” renegotiations.4 An agreement between two agents must evenly split the

surplus generated by their link over and above the expected payoffs they would receive in it’s

absence. This can be thought of as the transfers being robust to renegotiation if renegotiation

would result in the two agents benefitting equally from the link.5 From this principle we

obtain a recursive definition of how surpluses get divided on different networks. Applying the

axiomatization of the Myerson value provided in Myerson (1980), we show that the unique

division of surplus compatible with our requirements is the Myerson value,6 a network-specific

version of the Shapley value.7 The transfers required to implement the agreements we identify

3Although we consider a model in which there is perfect bilateral risk sharing, we could easily extend themodel so that some income is perfectly observed, some income is private, and there is perfect risk sharing ofobservable income and no risk sharing of unobservable income. This would be consistent with the theoreticalpredictions of Cole and Kocherlakota (2001) and the empirical findings of Kinnan (2011). In the CARAutilities setting, such unobserved income outside the scope of the risk-sharing arrangement does not affect ourresults.4Stole and Zwiebel (1996) model bargaining between many employees and an employer. This scenario can berepresented by a star network with the employer at the center.5The process is decentralized, as it envisages pairwise renegotiations. The result in Stole and Zwiebel (1996)that we extend is Theorem 1. Their Theorem 2 can also be extended to our setting, and this would providefully noncooperative foundations. Indeed, related noncooperative foundations are provided by Fontenay andGans (2013), while Navarro and Perea (2013) take a different approach to microfounding the Myerson value.Slikker (2007) also provides noncooperative foundations, although the game analyzed is not decentralized:offers are made at the coalitional level.6The Myerson value is defined for a transferable utility environment. However, agreements are made be-fore income realizations and determine expected utilities, and in our CARA setting, expected utilities aretransferable through state-independent transfers.7For investigations of the division of surplus in social networks in other contexts, see Calvo-Armengol (2001,2003), Corominas-Bosch (2004), Manea (2011), and Kets et al. (2011).


are particularly simple. Each agent receives an equal share of aggregate realized income (as

in Bramoulle and Kranton, 2007a) and then state independent transfers are made.

Although we derive the prediction of agents sharing the surplus according to the Myerson

value using a decentralized approach, such a split of surplus might also be appropriate when

negotiations are centralized. In such situations the Myerson value has normative appeal

on the grounds of fairness: two agents benefit equally from an investment they make, and

receive benefits proportional to their average marginal contributions to total surplus (from

establishing costly links).8

A key implication of the Myerson value determining the division of surplus is that agents

who are more centrally located, in a certain sense, receive a higher share of the surplus. More-

over, in our risk-sharing context it implies that agents receive larger payoffs from providing

“bridging links” to otherwise socially distant agents than from providing local connections.9

Empirical evidence supports this feature of our model—see Goyal and Vega-Redondo (2007),

and references therein from the organizational literature: Burt (1992), Podolny and Baron

(1997), Ahuja (2000), and Mehra et al. (2001).

In the network formation stage, we study the set of pairwise-stable networks (Jackson and

Wolinsky, 1996).10

Our general analysis considers a community comprised of different groups where all agents

within each group are ex-ante identical, and establishing links within groups is cheaper than

across groups. We also assume that the income realizations of agents within groups are

more positively correlated than across groups. Groups can represent different ethnic groups

or castes in a given village, or different villages. We find that there can be overinvestment

within groups but not underinvestment, whereas across groups underinvestment is likely to

be the main concern.

To see the intuition about overinvestment within groups, we first consider the case of

homogeneous agents, that is, when there is only one group. Using the inclusion–exclusion

principle from combinatorics,11 we develop a new metric to describe how far apart two agents

located in a network are, which we call the Myerson distance. Using this distance we provide

a complete characterization of stable networks. We show that for homogenous agents there

can never be underinvestment in social connections, as agents establishing an essential link

(connecting two otherwise unconnected components of the network) always receive a ben-

efit exactly equal to the social value of the link. However, overinvestment, in the form of

redundant links, is possible, and becomes widespread as the cost of link formation decreases.

8These motivations make the Myerson value a commonly used concept in the network formation literature.See a related discussion on pp. 422–425 of Jackson (2010).9More precisely, in Section 4 we introduce the concept of Myerson distance to capture the social distancebetween agents in the network, and show that a pair of agents’ payoffs from forming a relationship areincreasing in this measure.10Results from Calvo-Armengol and Ilkilic (2009) imply that under some parameter restrictions—for examplewhen agents are ex ante identical—the set of pairwise-stable outcomes is equivalent to the (in general morerestrictive) set of pairwise Nash equilibrium outcomes.11See Chapter 10 in van Lint and Wilson (2001).


Even though agents are ex-ante identical, if stable networks are asymmetric, inequality will

result. We identify a novel trade-off between efficiency and inequality. Among all possible

efficient network structures, we find that the most stable (in the sense of being stable for

the largest set of parameter values) results in the most unequal division of surplus (for any

inequality measure in the Atkinson class). Conversely, the least stable efficient network

entails the most equal division of surplus among all efficient networks. Although agents are

ex-ante identical, efficiency considerations push the structure of social connections towards

asymmetric outcomes that elevate certain individuals. Socially central individuals emerge

endogenously from risk-sharing considerations alone.

Turning attention to the case of multiple groups, we find that across-group underinvestment

becomes an issue when the cost of maintaining links across groups is sufficiently high.12 The

reason is that the agents who establish the first connection across groups receive less than the

social surplus generated by the link, providing positive externalities for peers in their groups.

To consider which agents are best incentivised to provide across group links we introduce a

new measure of network centrality which we term Myerson centrality. Agents more central in

this sense have better incentives to provide across group links. This provides a second force

pushing some agents within a group to be more central than others. For example, with two

groups, we show that the most stable efficient network structure involves stars within groups,

connected by their centers. This reinforces the trade-off between efficiency and equality in

the many-groups context.

Our analysis is premised on the basic assumption that social networks are endogenous

and that their structure affects how the surplus from risk sharing is split. There is growing

empirical evidence that risk-sharing networks respond to financial incentives, and that in

general risk-sharing networks form endogenously, in a way that depends on the economic

environment.13

We also provide some supporting empirical evidence to our theoretical findings, using

network data from 75 Indian villages. Because of data limitations (we do not observe income

realizations and actual transfers), our investigation is limited to how various factors correlate

with the structure of risk-sharing networks. We split the villagers into two groups, by caste.14

The theoretical analysis predicts that risk-sharing links are most valuable when they bridge

otherwise unconnected components. And when a link does not provide such a bridge, its

value depends on how far apart, suitably defined (which we call their Myerson distance),

the agents would otherwise be on the social network. We observe that as income variability

increases or within-caste incomes become less correlated, all else equal, the value of a risk-

sharing link between these villagers increases; the Myerson distances that can be observed in

12While across-group overinvestment remains possible, the main concern when across-group link costs arerelatively high is underinvestment.13See recent work by Binzel et al. (2014) and Banerjee et al. (2014b,c), which in different contexts look at howsocial networks respond to the introduction of financial instruments such as savings vehicles or microfinance.14There is an extensive literature that examines caste as a main social unit where risk sharing takes place(Townsend (1994), Munshi and Rosenzweig (2009), and Mazzocco and Saini (2012)).


a stable network therefore decrease, which in turn implies the following: (i) income variability

is positively associated with lower Myerson distances between unconnected agents, and (ii)

within-caste income correlation is associated with higher Myerson distances. The theory also

predicts that villagers have to be sufficiently central within their own caste (the threshold

depending again on income variability and within- versus across-caste income correlation) to

be incentivized to provide a risk-sharing link across castes. This yields two more predictions:

(iii) in villages with more income variability, more agents will have sufficient incentives to form

an across-caste link, and so the association between within-caste centrality and who provides

across-caste links is weaker, and (iv) in villages with more within-caste income correlation

relative to across-caste income correlation, more agents will again have sufficient incentives to

form across-caste links, and so the association between within-caste centrality and having an

across-caste link will again be weaker. In our analysis, we utilize that we not only have nearly

complete financial network data in every village, but we have nearly complete friendship

network data as well. As our theory pertains only to the financial network, we are able to

ask whether the patterns in the data that are consistent with our theory differentially play

out for the financial network instead of the social network. Our empirical findings support

the theoretical predictions, suggesting that even though our model is stylized, its qualitative

predictions can be captured in real world data. In particular they suggest that bargaining

considerations influence the structure of risk-sharing networks.

Ultimately, our empirical approach allows us to be more conservative than similar studies

in the literature (e.g., Karlan et al. (2009), Ambrus et al. (2014), and Kinnan and Townsend

(2012)). While the studies above have access to just a few of networks (e.g., 2, 1 and 16,

respectively), we have 75 networks and also multigraph data. Most studies, therefore, are

forced to do statistical inference within networks, which limits the number degree of correlated

shocks they are able to handle. Relative to this approach, our focus on the village level and

differencing out the social network is extremely conservative.

On the theory side, the studies on social networks and informal risk sharing that are most

related to ours include Bramoulle and Kranton (2007a,b), Bloch et al. (2008), Jackson et

al. (2012), Billand et al. (2012), Ali and Miller (2013a,b), and Ambrus et al. (2014). Many

of these papers focus on the enforcement issues we abstract from and investigate how social

capital can be used to sustain cooperation for lower discount factors than would otherwise be

possible. We take a complementary approach and instead focus on the distribution of surplus

and the incentives this creates for social investments. One way of viewing our approach is

an assumption on the discount factor in a dynamic version of our model. As long as the

discount factor is high enough, our equilibrium agreements satisfy the necessary incentive

compatibility constraints to be able to be enforced in equilibrium of a dynamic game.


Among the aforementioned papers, Bramoulle and Kranton (2007a,b) and Billand et al.

(2012) investigate costly network formation. Bramoulle and Kranton’s (2007a,b) model as-

sumes that the surplus on a connected income component is equally distributed, indepen-

dently of the network structure. This rules out the possibility of overinvestment or inequality,

and leads to different types of stable networks than in our model. Instead of assuming op-

timal risk-sharing arrangements, Billand et al. (2012) assume an exogenously given social

norm, which prescribes that high-income agents transfer a fixed amount of resources to all

low-income neighbors. This again leads to very different predictions regarding the types of

networks that form in equilibrium.

More generally, network formation problems are important. Establishing and maintaining

social connections (relationships) is costly, in terms of time and other resources. However,

on top of direct consumption utility, such links can yield many economic benefits. Papers

studying formation in different contexts include Jackson and Wolinsky (1996), Bala and

Goyal (2000), Kranton and Minehart (2001), Hojman and Szeidl (2008), and Elliott (2014).

Notable in this literature is a lack of empirical work, which can be attributed to a number

of innate difficulties that taking these models to data presents. Our many observations of

social networks that are relatively independent of each other, coupled with our approach to

circumventing data limitations, allow us to provide a first step towards testing predictions

based on the overall network structure. And although we study a specific network formation

problem tailored to risk sharing in villages, the general structure of our problem is relevant

to other applications.15

The remainder of the paper is organized as follows. Section 2 describes risk sharing on

a fixed network. In Section 3 we introduce a game of network formation with costly link

formation. We focus on network formation within a single group in Section 4 and then turn

to the formation of across-group links in Section 5. We then generate comparative static

results in Section 6, which we use to take our model to the data in Section 7. Section 8

concludes.

2. Preliminaries and Risk Sharing on a Fixed Network

To study social investments and the network formation problem, first we need to specify

what risk-sharing arrangements take place once the network is formed. Below we introduce

an economy in which agents face random income realizations, introduce some basic network

terminology, and discuss risk-sharing arrangements for a given network.

2.1. The socio-economic environment. We denote the set of agents in our model by N,

and assume that they are partitioned into a set of groups M. We let G : N → M be a

15For a different and more specific application, suppose researchers can collaborate on a project. Each re-searcher brings something heterogenous and positive to the value of the collaboration, so that the value ofthe collaboration is increasing in the set of agents involved. Collaboration is possible only when it takes placeamong agents who are directly connected to another collaborator and surplus is split according to the Myersonvalue (as in our work, motivated by robustness to renegotiations). Such a setting fits into our framework.


function that assigns each agent to a group; i.e., if G(i) = g then agent i is in group g.

One interpretation of the group partitioning is that N represents individuals in a village,

and the groups correspond to different castes. Another possible interpretation is that N

represents individuals in a larger geographic region (such as a district or subdistrict), and

groups correspond to different villages in the region.

Agents in N face uncertain income realizations. For tractability, we assume that incomes

are jointly normally distributed, with expected value µ and variance σ2 for each agent.16 We

assume that the correlation coefficient between the incomes of any two agents within the

same group is ρw, while between the incomes of any two agents not in the same group it is

ρa < ρw.17 That is, we assume that incomes are more positively correlated within groups

than across groups, so that all else equal, social connections across groups have a higher

potential for risk sharing.

Although we introduce the possibility of correlated incomes in a fairly stylized way, our

paper is one of the first to permit differently correlated incomes between different pairs of

agents. Such correlations are central to the effectiveness of risk-sharing arrangements, as

shown below.

We refer to possible realizations of the vector of incomes as states, and denote a generic

state by ω. We let yi(ω) denote the income realization of agent i in state ω.

Agents can redistribute realized incomes; hence their consumption levels can differ from

their realized incomes. We assume that all agents have constant absolute risk aversion

(CARA) utility functions:

(1) v(ci) = − 1

λe−λci ,

where ci is agent i’s consumption and λ > 0 is the coefficient of absolute risk aversion. The

assumption of CARA utilities, together with jointly normally distributed incomes, greatly

enhances the tractability of our model: as we show below it leads to a transferable utility

environment in which the implemented risk-sharing arrangements are relatively simple. This

utility formulation can also be considered a theoretical benchmark case with no income effects.

The empirical relevance of predictions from this benchmark model is examined in Section 6.

2.2. Basic network terminology. Before proceeding, we introduce some standard termi-

nology from network theory. A social network L is an undirected graph, with nodes N

corresponding to the different agents, and links representing social connections. Abusing

notation we also let L denote the set of links in the network. We will refer to the agents

linked to agent i, N(i;L) := j : lij ∈ L ⊂ N, as i’s neighbors. The degree centrality of an

agent is simply the number of neighbors she has (i.e., the cardinality of N(i;L)). An agent’s

16This specification implies that we cannot impose a lower bound on the set of feasible consumption levels.As we show below, our framework readily generalizes to arbitrary income distributions, but the assumptionof normally distributed shocks simplifies the analysis considerably.17It is well-known that for a vector of random variables, not all combinations of correlations are possible. Weimplicitly assume that our parameters are such that the resulting correlation matrix is positive semidefinite.


neighbors can be partitioned according to the groups they belong to. Let Ng(i;L) be i’s

neighbors on network L from group g. A path is a sequence of agents i, k, k′, . . . , k′′, j such

that every pair of adjacent agents in the sequence is linked. A minimal path between i and

j is a path where i is the first agent in the sequence, j is the last and there is no other path

between i and j that is a subsequence of that path. The path length of a path is the number

of agents in the path.

We will sometimes refer to subsets of agents S ⊆ N and denote the subgraphs they generate

by L(S) := lij ∈ L : i, j ∈ S. A subset of agents S ⊆ N is path connected on L if, for each

i ∈ S and each j ∈ S, there exists a path connecting i and j. For any network there is a

unique partition of N such that there are no links between agents in different partitions but

all agents within a partition are path connected. We refer to the cells of this partition as

network components. A shortest path between two path-connected agents i and j is a path

connecting i and j with a lower path length than any other. The diameter of a network

component C ⊂ L is d(C), the maximum value—taken over all pairs of agents in C—of the

length of a shortest path. A network component is a tree when there is a unique minimal

path between any two agents in the component. A line network is the unique (tree) network,

up to a relabeling of agents, in which there is a path from one (end) agent to the other (end)

agent that passes through all other agents exactly once. A star network is the unique tree

network, up to a relabeling of agents, in which one (center) agent is connected to all other

agents.

2.3. Risk-Sharing Agreements. We assume that income cannot be directly shared be-

tween agents i, j ∈ N unless they are connected, i.e., lij ∈ L. However, we let the state ω

(and hence every agent’s income realization) be publicly observed so agents can make trans-

fer arrangements contingent on it. We consider this environment with perfectly observable

incomes as a benchmark model, which is a relatively good description of village societies in

which people closely monitor each other. It is also straightforward to extend the model so

that some income is publicly observed (and shared) while the remaining income is privately

observed (and never shared). Results are very similar for this more general setting.18

Formally, a risk-sharing agreement on a network L specifies transfer tij(ω,L) = −tji(ω,L)

between neighboring agents i and j for every possible state ω. Abusing notation where there

should be no confusion we sometimes drop the second argument and write tij(ω) instead

of tij(ω,L). The interpretation is that in state ω agent i is supposed to transfer tij(ω)

units of consumption to agent j if tij(ω) > 0, and receives this amount from agent j if

tij(ω) < 0. Given a transfer arrangement between neighboring agents, agent i’s consumption

in state ω is ci(ω) = yi(ω)−∑

j∈N(i) tij(ω). It is straightforward to show that state-contingent

consumption plans (ci(·))i∈N are feasible, that is they can be achieved by bilateral transfers

18Kinnan (2011) finds evidence that hidden income can explain imperfect risk sharing in Thai villages relativeto the enforceability and moral hazard problems we are abstracting from. Cole and Kocherlakota (2001) showthat when individuals can privately store income, state-contingent transfers are not possible and risk sharingis limited to borrowing and lending.


between neighboring agents, if and only if∑

i∈C ci(ω) =∑

i∈C yi(ω) for every state ω and

network component C.

A basic assumption we make in our model is that given all other risk-sharing arrangements,

an agreement reached by linked agents i and j must leave no gains from trade on the table.

There must be no other agreement that can make both i and j strictly better off holding

fixed the agreements of other players. We call such transfers pairwise efficient.19

By the well-known Borch rule (see Borch (1962), Wilson (1968)) a necessary and sufficient

condition for this property is that for all neighboring agents i and j,

(2)

(∂vi(ci(ω))

∂ci(ω)

)/(∂vj(cj(ω))

∂cj(ω)

)=

(∂vi(ci(ω

′))

∂ci(ω′)

)/(∂vj(cj(ω

′))

∂cj(ω′)

)for every pair of states ω and ω′. But if this holds for all neighboring agents i and j then

the same condition must hold for all pairs of agents on a component of L, independently of

whether they are directly or indirectly connected. The results in Borch (1962) and Wilson

(1968) imply that the risk-sharing agreement is Pareto-efficient at the component level. For

this reason, below we establish some important properties of Pareto-efficient risk-sharing

arrangements on components.

Proposition 1 shows that the CARA utilities framework has the convenient property that

expected utilities are transferable, in the sense defined by Bergstrom and Varian (1985). This

can be used to show that ex-ante Pareto efficiency is equivalent to minimizing the sum of the

variances, and it is achieved by agreements that at every state split the sum of the incomes

on each network component equally among the members and then adjust these shares by

state-independent transfers. The latter determine the division of the surplus created by the

risk sharing agreement. We emphasize that this result does not require any assumption on

the distribution of incomes, only that agents have CARA utilities.

Proposition 1. For CARA utility functions certainty-equivalent units of consumption are

transferable across agents, and if L(S) is a network component, the Pareto frontier of ex-

ante risk-sharing agreements among agents in S is represented by a simplex in the space of

certainty-equivalent consumption. The ex-ante Pareto-efficient risk-sharing agreements for

agents in S are those that satisfy

min∑i∈S

Var(ci) subject to∑i∈Sci(ω) =

∑i∈Syi(ω) for every state ω,

and they are comprised of agreements of the form

ci(ω) =1

|S|∑k∈S

yk(ω) + τi for every i ∈ S and state ω,

where τi ∈ R is a state independent transfer.

19More formally, transfers tij(ω,L)ω∈Ω, ij:lij∈L are pairwise efficient for a network L if there is no pair of

agents ij : lij ∈ L and no alternative transfers t′ij(ω,L)ω∈Ω, ij:lij∈L such that t′kl(ω,L) = tkl(ω,L) for allkl 6= ij and all ω ∈ Ω, that gives both i and j strictly higher expected utility.


The proof of Proposition 1 is in Appendix C. Proposition 1 implies that the total sur-

plus generated by efficient risk-sharing arrangements is an increasing function of the re-

duction in aggregate consumption variance (the sum of consumption variances). For a

general distribution of shocks, this function can be complicated. However, if shocks are

jointly normally distributed then ci = 1|S|∑

k∈S yk + τi is also normally distributed, and

E(v(ci)) = E(ci) − λ2 Var(ci).

20 Hence in this case the total social surplus generated by

efficient risk-sharing agreements is proportional to the aggregate consumption variance re-

duction. This greatly simplifies the computation of surpluses in the analysis below.

We use TS(L) to denote the expected total surplus generated by an ex-ante Pareto-efficient

risk-sharing agreement on network L, relative to agents consuming in autarky:

TS(L) := CE(

∆ Var(L, ∅)),(3)

where, for L′ ⊂ L, ∆ Var(L,L′) is the additional variance reduction obtained by efficient

risk-sharing on network L instead of L′, and CE(·) denotes the certainty-equivalent value of

a variance reduction.

2.4. Division of Surplus. The assumption that neighboring agents make pairwise efficient

risk-sharing agreements pins down agreements up to state-independent transfers between

neighboring agents, but does not constrain the latter transfers (hence the division of surplus)

in any way. To determine these transfers, we follow the approach in Stole and Zwiebel (1996)

and require that agreements are robust to split the difference renegotiations. This implies

that the transfer is set in a way such that the incremental benefit that the link provides to the

two agents is split equally between them. The motivation for this as provided by Stole and

Zwiebel is to suppose that before agreements have been finalized any agent can opt out of an

agreement with a neighbor and then renegotiate it so that he benefits from it as much as his

neighbor.21 Of course, finding an agreement that i and j can reach without either wanting

to renegotiate it requires them splitting the incremental benefits equally. And determining

this split requires calculating the expected payoffs i and j would receive if they did not have

an agreement. We therefore have to consider what agreements would prevail on the network

without lij to find the risk sharing agreements i and j can reach on L, and so on. This results

in a recursive system of conditions.

More formally, for a network L a contingent transfer scheme

(4) T (L) := tij(ω,L′)ω∈Ω, L′⊆L, ij:lij∈L,

specifies all transfers made in all subnetworks of L in all states of the world. The expected

utility of agent i on a network L′ ⊆ L given a contingent transfer scheme T (L) is denoted

20See, for example, Arrow (1965).21For a detailed motivation of this assumption, and for noncooperative microfoundations, see Stole and Zwiebel(1996).


1 3 4 2

𝑁𝑇𝑆 = 3𝑉

(a) Line network

2 4 3

𝑁𝑇𝑆 = 2𝑉

(b) Three nodes

4 3

𝑁𝑇𝑆 = 𝑉

(c) Two nodes

Figure 1. To find (gross) expected utilities that are robust to split the dif-ference renegotiations on the (formed) line network shown we need to considerthe expected utilities that would be obtained on all subnetworks.

ui(L′, T (L)). Where there should be no confusion, we will abuse notation and drop the

second argument.

For any network L, the expected utility vector (u1, ..., u|N|) is robust to split the difference

renegotiation if there is a contingent transfer scheme T (L) such that ui = ui(L, T (L)) for

every i ∈ N and the following conditions hold:

(i) ui(L′)− ui(L′ \ lij) = uj(L

′)− uj(L′ \ lij) for every lij ∈ L′ and L′ ⊆ L;

(ii) transfers tij(ω,L′)ω∈Ω, ij:lij∈L′ are pairwise efficient for all L′ ⊆ L.

Suppose we want to find payoffs robust to split the difference renegotiation for the line

network shown in Figure 1a. A first necessary condition is that agents 1 and 2 benefit equally

from their link so that u1(L)−u1(L\l12) = u2(L)−u2(L\l12). But in order to ensure this

condition is satisfied, we need to know u1(L \ l12) and u2(L \ l12). Without the link l12

agent 1 is isolated so u1(L \ l12) = 0. However, to find u2(L \ l12) we need to find payoffs

for the three node network in Figure 1b. For this network robustness to split the difference

renegotiation requires that u2(L\l12)−u2(L\l12, l23) = u3(L\l12)−u3(L\l12, l23).While u2(L \ l23, l23, ) = 0, we need to consider the two node network shown in Figure 1c

to find u3(L \ l12, l23). For this network, payoffs must satisfy u3(L \ l12, l23) − u3(L \l12, l23, l34) = u4(L \ l12, l23) − u4(L \ l12, l23, l34). As u3(L \ l12, l23, l34) = u4(L \l12, l23, l34) = 0, the above condition simplifies to u3(L\l12, l23) = u4(L\l12, l23) = V/2,

where the last equality follows from pairwise efficiency. Considering the three node network

again, we now have the condition u2(L\l12) = u3(L\l12)−V/2. As the link l23 generates

an incremental surplus of V to be split between agents 2 and 3, pairwise efficiency implies

that u2(L\l12) = V/2 and u3(L\l12) = V . Finally, returning to the line network, we now

have u1(L) = u2(L)− V/2. As the link l12 generates incremental surplus of V , u1(L) = V/2

and u2(L) = V .22

22This argument only outlines why the payoffs u1(L) = V/2 and u2(L) = V are necessary for robustness tosplit the difference renegotiations. By considering all other subnetworks, it can be shown that the payoffsu1(L) = u4(L) = V/2 and u2(L) = u3(L) = V are the unique payoffs that are robust to split the differencerenegotiations.


Below we show that the requirement of robustness to split the difference renegotiation

implies that the total surplus created by the risk-sharing agreement is divided among agents

according to the Myerson value (Myerson 1977, 1980). The Myerson value is a cooperative

solution concept defined in transferable utility environments that is a network-specific version

of the Shapley value. The basic idea behind it is the same as for the Shapley value. For any

order of arrivals of the players, the incremental contribution of an agent i to the total surplus

can be derived as the difference between the total surpluses generated by the subgraph of

L(S) the subgraph L(S \ i) if agents S \ i arrive before i. It is easy to see that, for any

arrival order, the total surplus generated by L gets exactly allocated to the set of all agents.

The Myerson value then allocates the average incremental contribution of a player to the

total surplus, taken over all possible orders of arrivals (permutations) of the players, as the

player’s share of the total surplus. Thus, agent i’s Myerson value is23

(5) MVi(L) :=∑S⊆N

(|S| − 1)!(|N| − |S|)!|N|!

(TS(L(S))− TS(L(S \ i))

).

Proposition 2. For any network L, any risk-sharing agreement that is robust to split

the difference renegotiation yields expected payoffs to agents equal to their Myerson values:

ui(L) = MVi(L).

Proof. Theorem 1 of Myerson (1980) states that there is a unique rule for allocating surplus

for all subnetworks of L that satisfies the requirements of efficiency at the component level

(note that this is an implicit requirement in Myerson’s definition of an allocation rule) and,

what Myerson (1980) defines as the equal-gains principle. Moreover, the expected payoff the

above rule allocates to any player i is MVi. Requirement (i) in our definition of robustness

to split the difference renegotiation is equivalent to the equal-gains principle as defined in

Myerson (1980). Theorem 1 of Wilson (1968) implies that efficiency at the component level

is equivalent to pairwise efficiency between neighboring agents, which is requirement (ii) in

our definition of robustness to split the difference renegotiation. The result then follows

immediately from Theorem 1 of Myerson (1980).

Proposition 2 is a direct implication of Myerson’s axiomatization of the value. A special

case of Proposition 2 is Theorem 1 of Stole and Zwiebel (1996), which in effect restricts

attention to a star network.24 Our contribution is to point out that their connection between

23Our assumption that there is perfect risk sharing among path-connected agents ensures that a coalition ofpath connected agents generates the same surplus regardless of the exact network structure connecting them.This means that we are in the communication game world originally envisaged by Myerson. We do not requirethe generalization of the Myerson value to network games proposed in Jackson and Wolinsky (1996), whichsomewhat confusingly is also commonly referred to as the Myerson value.24Relative to Myerson’s axiomatization, Stole and Zwiebel (1996) generate the key system of equations throughconsidering robustness to renegotiations as we describe above, while Myerson wrote down the system of equa-tions based only on fairness considerations. Stole and Zwiebel (1996) also provide non-cooperative bargainingfoundations that underpin this system.


robustness to split the difference renegotiations and the Shapley value can be extended to

apply to all networks.

The above result shows that any decentralized negotiation procedure between neighboring

agents that satisfies two natural properties (not leaving surplus on the table, and robustness

to split the difference negotiations) leads to the total surplus created by risk-sharing divided

according to the Myerson value, and to state-independent transfers between neighboring

agents that implement this surplus division. Hence, from now on we assume that in the

network formation process, all agents expect the surplus to be divided according to the

Myerson value implied by the network that eventually forms.

Although we followed a decentralized approach to get to the implication that surplus is

divided by the Myerson value, we note that on normative grounds such a division is also cogent

in contexts in which there is a centralized community level negotiation over the division of

surplus. This is because the Myerson value is a formal way of defining the fair share of an

individual from the social surplus, as his average incremental contribution to the total social

surplus (where the average is taken across all possible orders of arrival of different players,

in the spirit of the Shapley value).

3. Investing in Social Relationships

Having defined how formed networks map into risk-sharing arrangements, we can now

consider agents’ incentives to make social investments. We begin by providing the overall

framework for the analysis. Then we look at a special case of our model, in which there is a

single group. Building on these results we then consider the multiple group case.

In this section we formalize a game of network formation in which establishing links is

costly, define efficient networks and identify different types of investment inefficiency.

We consider a two-period model in which in period 1 all agents simultaneously choose

which other agents they would like to form links with, and in period 2 agents agree upon

the ex-ante Pareto-efficient risk-sharing agreement specified in the previous section (i.e., the

total surplus from risk sharing is distributed according to the Myerson value), for the network

formed in the first period.25 Implicit in our formulation of the timing of the game is the view

that relationships are formed over a longer time horizon than that in which agreements are

reached about risk sharing. By the time such agreements are being negotiated, the network

structure is fixed, and investments into forming social relationships are sunk.

Formally, in period 1 we consider a network formation game along the lines of Myerson

(1991): all agents simultaneously choose a subset of the other agents, indicating who they

would like to form links (relationships) with. A link is formed between two agents if and only

if they both want to form it (i.e., if both agents select each other). When agent i forms a

25For a complementary treatment of network formation when surplus is split according to the Myerson value,see Pin (2011).


link, he pays a cost κw > 0 if the link is with someone in the same group and κa > κw if the

link is with someone from a different group.

The collection of links formed in period 1 becomes social network L. Normalizing the utility

from autarky to 0, we abuse notation26 and let agent i’s net expected utility if network L

forms be

(6) ui(L) = MVi(L)− |NG(i)(i;L)|κw −(|N(i;L)| − |NG(i)(i;L)|

)κa.

The solution concept we apply to the simultaneous-move game described above is pairwise

stability. A network L is pairwise stable with respect to expected utilities ui(L)i∈N if and

only if for all i, j ∈ N, (i) if lij ∈ L then ui(L)−ui(L\lij) ≥ 0 and uj(L)−uj(L\lij) ≥ 0;

and (ii) if lij /∈ L then ui(L ∪ lij) − ui(L) > 0 implies uj(L ∪ lij) − uj(L) < 0. In words,

pairwise stability requires that no two players can both strictly benefit by establishing an

extra link with each other, and no player can benefit by unilaterally deleting one of his links.

From now on we will use the terms pairwise-stable and stable interchangeably.

Existence of a pairwise-stable network in our model follows from a result in Jackson (2003),

stating that whenever payoffs in a simultaneous-move network formation game are determined

based on the Myerson value, there exists a pairwise-stable network.

Our specification assumes that two agents forming a link have to pay the same cost for

establishing the link. However, the set of stable networks would remain unchanged if we

allowed the agents to share the total costs of establishing a link arbitrarily.27 This is because

for any link, the Myerson value rewards the two agents establishing the link symmetrically.

Hence the agents can find a split of the link-formation cost such that establishing the link is

profitable for both of them if and only if it is profitable for both of them to form the link with

an equal split of the cost. For this reason we stick with the simpler model with exogenously

given costs.

A network L is efficient when there is no other network L′—and no risk sharing agreement

on L′—that can make everyone at least as well off as they were on L and someone strictly

better off. Let |Lw| be the number of within-group links, and let |La| be the number of

across-group links. As expected utility is transferable in certainty-equivalent units, efficient

networks must maximize the net total surplus NTS(L):

NTS(L) := TS(L)− 2|Lw|κw − 2|La|κa,(7)

Clearly, two necessary conditions for a network to be efficient are that the removal of a

set of links does not increase NTS(L) and the addition of a set of links does not increase

26In the previous section when investments had already been sunk we used ui(L) to denote i’s expected payoffbefore link formation costs.27More precisely, we could allow agents to propose a division of the costs of establishing each link as well asindicating who they would like to link to, and a link would then form only if both agents indicate each otherand they propose the same split of the cost. A network would then be stable if it is a Nash equilibrium of thisexpanded network formation game and if there is no new link lij 6∈ L , and some split of the cost of formingthis link, that would make both i and j strictly better off if formed.


NTS(L). If there exists a set of links the removal of which increases NTS(L), we will

say there is overinvestment inefficiency. If there exists a set of links the addition of which

increases NTS(L), we will say there is underinvestment inefficiency.28 A network is robust to

underinvestment if there is no underinvestment inefficiency and no agent can strictly benefit

from deleting a link would result in underinvestment inefficiency. A network is robust to

overinvestment if there is no overinvestment inefficiency and no pair of agent i, j can both

strictly benefit from creating the link lij .

We will say that a link lij is essential if after its removal i and j are no longer path

connected while it is superfluous if after its removal i and j are still path connected.

Remark 3. Preventing overinvestment requires that all links be essential. Superfluous links

create no social surplus and are costly. In all efficient networks, therefore, every component

must be a tree.

In most of the analysis below, we focus on investigating the relationship between stable

networks and efficient networks. Additionally, we investigate the amount of inequality pre-

vailing in equilibria in our model. For this, we will use the Atkinson class of inequality

measures (Atkinson, 1970). Specifically we consider a welfare function W : R|N| → R that

maps a profile of expected utilities into the real line such that

(8) W (u) =∑i∈N

f(ui),

where f(·) is assumed to be an increasing, strictly concave and differentiable function. The

concavity of f(·) captures the social planner’s preference for more equal income distributions.

Supposing all agents instead received the same expected utility u′, we can pose the question

what aggregate expected utility is required to keep the level of the welfare function constant.29

In other words we find the scalar u′ : |N|f(u′) =∑

i∈N f(ui). Letting u = (1/|N|)∑

i∈N ui

be the mean expected utility, Atkinson’s inequality measure (or index) is given by

(9) I(f) = 1− u′

u∈ [0, 1].

We let I be the set (class) of Atkinson inequality measures and note that any I(f) ∈I equals zero if and only if all agents receive the same expected utility.30 Two different

inequality measures from the Atkinson class can rank the inequality of two distributions

28Note that these definitions are not mutually exclusive (there can be both underinvestment and overin-vestment inefficiency) or collectively exhaustive (inefficient networks can have neither underinvestment noroverinvestment inefficiency if an increase in the net total surplus is only possible by the simultaneous additionand removal of edges).29This exercise is analogous to the certainty equivalent exercise that can be undertaken for an agent facingstochastic consumption.30As f(·) approaches the linear function the social planner cares less about inequality and I(f) → 0. Never-theless, strict concavity prevents I(f) equaling 0 unless all agents receive the same expected utility.


differently. However, certain pairs of distributions are ranked the same way by all members

of the class, such as when one distribution is a mean-preserving spread of the other one.

4. Within-Group Network Formation

In this section we assume that |M| = 1, that is, that agents are ex-ante symmetric, and

any differences in their outcomes stem from their stable positions on the social network. This

will lay the foundations for the more general case considered in the next section.

We begin our investigation by proving a general characterization of the set of stable net-

works. Recall that a minimal path between i and j is any path between i and j such that

no other path between i and j is a subsequence of that path. If there are K minimal paths

between i and j on the network L, we let P(i, j, L) = P1(i, j, L), . . . , PK(i, j, L) be the set

of these paths. For every k ∈ 1, . . . ,K, let |Pk(i, j, L)| be the cardinality of the set of agents

on the minimal path Pk(i, j, L).31 We can now use these definitions to define a quantity that

captures how far away two agents are on a network in terms of the probability that for a

random arrival order they will be connected without a direct link when the second of the two

agents arrives. We will refer to this distance as the agents’ Myerson distance:

(10) md(i, j, L) :=1

2−|P(i,j,L)|∑k=1

(−1)k+1

∑1≤i1<···<ik≤|P(i,j,L)|

(1

|Pi1 ∪ · · · ∪ Pik |

) .

This expression calculates the probability that for a random arrival order the link lij will

be essential immediately after i arrives,32 using the classic inclusion–exclusion principle from

combinatorics. This probability is important because it affects i’s incentives to link to j.

As an illustration, suppose that there is a unique indirect path P1(i, j, L) between i and

j that contains K agents, including i and j. We then have md(i, j, L) = 1/2 − 1/K. To

see where this expression comes from, note that there are two reasons why lij might not be

essential when i arrives. First, j might not yet have arrived. This occurs with probability

1/2. Second, all other agents on the path P1(i, j, L), including j, might have already arrived.

This occurs with probability 1/K. The probability of both events occurring is 0 because

collectively they require j to be both present and absent, so we can just sum them. Thus,

the probability that lij is essential when i arrives, is 1− 1/2− 1/K = md(i, j, L).

Suppose now that there are two (minimal) paths between i and j, P1(i, j, L) and P2(i, j, L),

on the network L. Suppose that P1(i, j, L) = i, i′, i′′, j and P2(i, j, L) = i, i′, i′′′, i′′′′, j.We need to find the probability that all the agents on at least one of these paths are present

when i arrives. To avoid double counting, we need to add the probability that all the agents

on P1(i, j, L) are present (1/4) to the probability all the agents on P2(i, j, L) are present (1/5)

31For example, for a path Pk(i, j, L) = i, i′, i′′, j, |Pk(i, j, L)| = 4 and for a path Pk′(i, j, L) = i, i′, i′′′, i′′′′, j,|Pk′(i, j, L)| = 5. Finally, we we will let |Pk(i, j, L) ∪ Pk′(i, j, L)| = 5 denote number of different agents onpath Pk(i, j, L) or path Pk′(i, j, L).32If for a given arrival order, agents S ⊆ N arrive before i, then lij is essential immediately after i arrives ifit is essential on the network L(S ∪ i).


and then subtract the probability that all the agents on both paths are present (1/6).33 So

md(i, j, L) = 1−1/2−1/4−1/5+1/6. The Myerson distance calculation provides the general

way of accounting for the probability that all the agents on at least one of the possible paths

are present.

Lemma 4. If all agents are from the same group network L is pairwise stable if and only if

(i) md(i, j, L \ lij) ≥ κw/V for all lij ∈ L, and

(ii) md(i, j, L) ≤ κw/V for all lij 6∈ L.

The proof is relegated to Appendix C. Recall from equation 3 that the social benefits of

a link is proportional to the variance reduction it generates. For a single group, if a link lij

is essential in the network L ∪ lij, then this variance reduction is ∆ Var(L ∪ lij, L) =

(1− ρw)σ2.

The crucial feature of this expression is that it does not depend on size of the network

components the link lij connected on L. Although in general the size of these components

does affect the consumption variance, two effects exactly offset each other.34 On the one

hand, in larger components there are more people to benefits from the essential link. On the

other hand, people are already able to smooth there consumption effectively.

As the social value of a non-essential, or superfluous link, is always zero the total surplus

generated by a network L takes a very simple form. Let f(L) be the number of network

components on L. Then TS(L) = CE(

∆ Var(L, ∅))

= (|N| − f(L))λ2 (1 − ρw)σ2. Since the

surplus created by any essential link is

(11) V :=λ

2(1− ρw)σ2,

the total gross surplus is equal to the latter constant times the number of network compo-

nent reductions obtained relative to the empty network.

To consider individual incentives to form links we can use the definition of the Myerson

value and consider the average marginal contribution an agent makes to total surplus over

all possible arrival orders. Specifically, we want to consider the increase in i’s Myerson value

due to a link lij . The link lij will reduce the number of components in the graph by one

when i arrives relative to the counterfactual component reduction without lij , if and only

if j has already arrived and there is no other path between i and j. In other words, the

link increases i’s marginal contribution to total surplus if and only if it is essential when

i is added. Moreover, for the permutations in which lij is essential it contributes V to i’s

33Note that all the agents on both P1(i, j, L) and P2(i, j, L) will be present if and only if i′, i′′, i′′′, i′′′′ and jare present before i arrives.34Let L(S1) and L(S2) be the network components of agent i and agent j on network L \ lij, and let|S1| = s1 and |S2| = s2. Then the sum of consumption variances on L(S1) and L(S2) (with Pareto-efficient

risk sharing) are s1+s1(s1−1)ρws1

σ2 and s2+s2(s2−1)ρws2

σ2, respectively. Once S1 and S2 are connected through

lij , the sum of consumption variances on L(S1 ∪ S2) becomes s1+s2+(s1+s2)(s1+s2−1)ρws1+s2

σ2. This implies that

the consumption variance reduction induced by the link lij is ∆ Var(L ∪ lij, L) = (1− ρw)σ2.


marginal contribution to total surplus. Averaging over arrival order, the value to i of the link

lij ∈ L is md(i, j, L\lij)V , while the value to establishing a new link lij 6∈ L is md(i, j, L)V .

If a link lij is essential on L then for any arrival order, there will always be a component

reduction of 1 when the later of i or j is added. Therefore, md(i, j, L) = 1/2, and lij will be

formed as long as V > 2κw. As V is the social value of forming the link and 2κw is the total

cost of forming it, when all agents are from the same group there is never underinvestment

in a stable network or overinvestment in an essential link.

Proposition 5. If all agents are from the same group then there is never underinvestment

in a stable network. Furthermore, there is never overinvestment in an essential link.

The proof is relegated to Appendix C. When all agents are from the same group Proposition

5 establishes that there is never overinvestment in an essential link, but overinvestment in

superfluous links is possible. If the costs of link formation are low enough then agents will

receive sufficient benefits from establishing superfluous links to be incentivized to do so. Even

if a link lij is superfluous on L, for some arrival orders it will be essential on the induced

subnetwork at the moment when i is added and make a positive marginal contribution to

total surplus.35 An example of such overinvestment is shown in Appendix B.

An immediate implication of Proposition 5 is that if all agents are from the same group

and 2κw > V then the only stable network is the empty one and this network is efficient,

while if 2κw < V then all stable networks have only one network component (all agents are

path connected). For the remainder of the paper we focus on the parameter range for which

the empty network is inefficient for a single group and assume 2κw < V . We refer to this as

our regularity condition and omit it from the statement of subsequent results.

Under this regularity condition the set of efficient networks are the set of tree networks

in which all agents are path connected. In other words, all agents must be in the same

component and all links must be essential. We will now focus on which, if any, of these

efficient networks are stable. As by Proposition 5 there is never any underinvestment in any

stable network the only reason an efficient network will not be stable is if two agents have

a profitable deviation by forming an additional (superfluous) link. We therefore focus on

investigating what network structures minimize incentives for overinvestment. As we will

see, this question is also related to the issue of inequality that different network structures

imply.

Figure 2 illustrates three networks: A line (Figure 2a); a circle (Figure 2b) and a star

(Figure 2c). While the line and star networks are efficient, the circle network is not as it

includes a superfluous link. Among the two efficient networks, the star is more stable than

the line. Applying Lemma 4, whenever the line is stable so is the star but there are parameter

values for which the star is stable and the line is not. While the star is more stable than

the line, it also results in more inequality. The expected uitlity distribution obtained on the

35Consider, for examples, arrival orders in which i arrives first and j arrives second.


1 3 4 2

𝑉 − 2κ𝑤2

(𝑉 − 2κ𝑤) 𝑉 − 2κ𝑤2

(𝑉 − 2κ𝑤)

(a) Line

1

3 4

2

3𝑉

4−κ𝑤

3𝑉

4−κ𝑤

3𝑉

4−κ𝑤

3𝑉

4−κ𝑤

(b) Circle

1

3 4

2

𝑉 − 2κ𝑤2

𝑉 − 2κ𝑤2

𝑉 − 2κ𝑤

2

3𝑉 − 2κ𝑤

2

(c) Star

Figure 2. Three possible network structures for connecting 4 agents and theresulting net payoffs.

line network can be generated from that obtained on the star network by the best off agent

(agent 2) transferring (V −2κw)/2 > 0 units of expected utility to one of the worst off agents

(agent 3). This is enough to ensure that the expected utility distribution on the star is more

unequal than the expected utility distribution on the line for any inequality measure in the

Atkinson class. We generalize these insights in Proposition 6.

Proposition 6. Suppose all agents are from the same group.

(i) If there exists an efficient stable network then star networks are stable, and for a

range of parameter specifications only star networks are stable. If a line network is

stable then all efficient networks are stable.

(ii) For all inequality measures in the Atkinson class, among the set of efficient network,

star networks and only star networks maximize inequality, while line networks and

only line networks minimize inequality.

The proof is in Appendix C but we provide some intuition after we discuss the result.

Proposition 6 states that, in a certain sense, among the set of efficient networks the star is

the most stable but maximizes inequality, while the line minimizes inequality but is least

stable. This indicates a novel tension between efficiency and inequality.

To gain intuition for Proposition 6, recall that Proposition 5 implies that an efficient

network will be stable if and only if no pair of players have a profitable deviation in which

they form a superfluous link. By Lemma 4 the incentives for two agents to form such a link

are strictly increasing in their Myerson distance. Thus, a network is stable if and only if

the pair of agents furthest apart from each other, in terms of their Myerson distance, cannot

benefit from forming a link. As efficient networks are tree networks, the Myerson distance

between any two agents depends only the length of the unique minimal path between them.36

The longest minimal path between any pair of agents is, by definition, the diameter of the

36Suppose d is the number of agents on the unique path connecting i and j. The probability that this pathexists when agent i arrives is 1/d. In addition, if agent j has not yet arrived, which occurs with probability1/2, i would not benefit from the link lij , so i’s expected payoff from forming a superfluous link to j is(1− 1/2− 1/d)V . We also note that as d gets large, this converges to V/2 which is the value i receives fromforming an essential link.


network d(L). So, an efficient network is stable if and only if its diameter is sufficiently small.

More precisely, an efficient network L is stable if and only if its diameter is weakly less than

d(κw, V ), where d(κw, V ) is increasing in κw, decreasing in V and integer valued.37

Let Le(N) be the set of efficient networks. Star networks have the smallest diameter among

networks within this set, while line networks have the largest diameter among networks within

this set. This establishes part (i) of Proposition 6.

To gain intuition for part (ii) a first step is noting that on any efficient network agents’

net payoffs are proportional to their degree (i.e., the number of neighbours they have):38

ui(L) = |N(i;L)|(V/2−κw). The key insight is then showing that for any network in the set

Le(N), the star network can be obtained by rewiring the network (deleting a link lij ∈ L and

adding a link lik 6∈ L) in such a way that at each step we increase the degree of the agent

who already has the highest degree and reduce the degree of some other agent and obtain a

new network in Le(N). This process transfers expected utility to the agent with the highest

expected payoff from some other agent. This increases inequality for any inequality measure

within the Atkinson class. Likewise, we can obtain the line network from any network in the

set Le(N) by rewiring the network to decrease the degree of the agent with the highest degree

at every step. This transfers expected utility from agent with the highest expected payoff to

some other agent and decreases inequality for any inequality measure in the Atkinson class.

To summarize, this section identifies two novel downsides to informal risk sharing agree-

ments. First, they promote a misallocation of villagers time towards excessive social capital

accumulation. Villagers are incentivized to form links with a view to becoming more central

within the risk sharing network in order to appropriate a larger share of the surplus gener-

ated by risk sharing. Secondly, even when investments into social capital are efficient the

networks that can be supported in equilibrium generate social inequality, and this translates

into (potentially severe) financial inequality.

5. Connections Across Groups

We now generalize our model by permitting multiple groups. These different groups might

correspond to people from different villages, different occupations, or different social status

groups, such as castes. We will first show that (under our regularity condition) there is still

never any underinvestment within a group. However, this does not apply to links that bridge

groups. As, by assumption, incomes are more correlated within a group than across a group,

there can be significant benefits from establishing such links and not all these benefits accrue

to the agents forming the link. Intuitively, an agent establishing a bridging link to another

group provides other members of his group with access to a less correlated income stream,

which benefits them. As agents providing such bridging links are unable to appropriate all

37We show in the proof that d(κw, V ) = b2V/(V − 2κw)c.38This is also know as an agent’s degree centrality.


the benefits these links generate, and these links are relatively costly to establish, there can

be underinvestment.

To analyze the incentives to form links within a group, we first need to consider the variance

reduction obtained by a within-group link. Such a link may now connect two otherwise

separate components comprised of arbitrary distributions of agents from different groups.

Suppose the agents in S0 ∪ · · · ∪Sk and the agents in S0 ∪ · · · ∪ Sk form two distinct network

components, where for every i ∈ 0, ..., k, the agents in Si and those in Si are all from group

i. Consider now a potential link lij connecting the two otherwise disconnected components.

Letting s0 be the number of agents in group 0, the variance reduction obtained is:39

(12) ∆ Var(L ∪ lij , L) =

(1− ρw) +

∑ki=0

(si∑k

j=0 sj − si∑k

j=0 sj

)2(∑ki=0 si

)(∑ki=0 si

)(∑ki=0 si + si

)(ρw − ρa)

σ2.

The key feature of this variance reduction is that it is always weakly greater than (1−ρw)σ2,

which is the variance reduction we found in the previous section when all agents were from

the same group. Thus, the presence of across-group links only increases the incentives for

within-group links to be formed. A within-group link can now give (indirect) access to less

correlated incomes from other groups and so is weakly more valuable. This implies that there

will still be no underinvestment under our regularity condition that 2κw < V .40 The above

reasoning is formalized by Proposition 7.

Proposition 7. There is no underinvestment between any two agents from the same group

in any stable network.

The proof of Proposition 7 is in Appendix C. While underinvestment is not possible within

group, it is possible across groups. An example of this is shown in Appendix B. Although

when all agents are from the same group the value of an essential link does not depend on

the sizes of the components it connects, the value of an essential link connecting two different

groups of agents increases in the sizes of the components. To demonstrate this formally,

consider an isolated group that has no across-group connections and consider the incentives

for a first such connection to be formed. Thus the first component consist of agents from

a single group, say group 0. We let the second component consist of agents from one or

more of the other groups (1 to k). The variance reduction obtained by connecting these two

components is

39By definition∆ Var(L ∪ lij , L) = Var(L(S0, ...,Sk)) + Var(L(S0, ..., Sk))−Var(L(S0 ∪ S0, ...,S ∪ Sk)).

Recalling that

Var(L(S0,S1, ...,Sk)) =

(k∑i=0

(si + si(si − 1)ρw) + 2ρa

k−1∑i=0

(si

k∑j=i+1

sj)

)σ2

/k∑i=0

si ,

some algebra yields the result.40Recall that this regularity condition just requires that it is efficient for two agents in the same group, bothwithout any other connections, to form a link.


(13) ∆ Var(L ∪ lij , L) =

(1− ρw) +

s0

((∑ki=1 si

)2+∑k

i=1 s2i

)(∑k

i=1 si

)(s0 +

∑ki=1 si

) (ρw − ρa)

σ2,

which is increasing in s0:

(14)∂∆ Var(L ∪ lij , L)

∂s0=

(∑ki=1 si

)2+∑k

i=1 s2i(

s0 +∑k

i=1 si

)2 (ρw − ρa)σ2 > 0.

The inequality follows since ρw > ρa. Thus if agents i and j who connect two otherwise

unconnected groups they receive a strictly smaller combined private benefit than the social

value of the link. To see why, suppose that on the network L the link lij is essential, and

without lij there would be two components, the first connecting agents from group G(i)

and the second connecting agents from group G(j) 6= G(i). Consider the Myerson value

calculation. For arrival orders in which i or j is last to arrive, the value of the additional

variance reduction due to lij obtained upon the arrival of the later of i or j, is the same as

its marginal social value, i.e., the value of variance reduction obtained by lij on L. For any

other arrival order the value of variance reduction due to lij when the later of i or j arrives

is strictly less. Averaging over these arrival orders, the link lij contributes less to i and j’s

combined Myerson values than its social value, leading to the possibility of underinvestment.

Besides underinvestment, overinvestment is also possible across groups. Forming super-

fluous links will increase an agent’s share of surplus without improving overall risk sharing

and can therefore create incentives to overinvest. Nevertheless, when κa is relatively high,

underinvestment rather than overinvestment in across-group links will be the main efficiency

concern. In many settings, within-group links are relatively cheap to establish in comparison

to across-group links. For example, when the different groups correspond to different castes,

as in our data, it can be quite costly to be seen interacting with members of the other caste

(e.g., Srinivas (1962), Banerjee et al. (2013b)). Motivated by this, and because across-group

links are considerably sparser in our data (to be described in the next section) than within-

group links, we focus our attention on this parameter region. More concretely, below we

investigate what within-group network structures create the best incentives to form across-

group links and what network structures minimize the incentives for overinvestment within

group. Remarkably, we find that these two forces push within-group network structures in

the same direction, and in both cases towards inequality in the society.

We begin by considering within-group overinvestment, which corresponds to the formation

of superfluous links within-group links. We found in the previous section that when all agents

are from the same group the star is the efficient network that minimized the incentives

for overinvestment. However, once we include links to other groups, the analysis is more

complicated. The variance reduction a within-group link generates is still 0 if the link is


superfluous, but when the link is essential it depends on the distribution of agents across

the different groups the link grants access to. Moreover, the variance reduction may be

decreasing or increasing in the numbers of people in those groups.41 This makes the Myerson

value calculation substantially more complicated. When all agents were from the same group

all that mattered was whether the link was essential when added. Now, for each arrival order

in which the link is essential, we also need to keep track of the distribution of agents across

the different groups that are being connected. Nevertheless, our earlier result generalizes to

this setting, although the argument establishing the result is more subtle.

To state the result, it is helpful to define a new network structure. A center-connected star

network is a network in which all within-group network structures are stars and all across-

group links are held by the center agents in these stars. We denote the set of center-connected

star networks by LCCS .

Proposition 8. If any efficient network L is robust to overinvestment within group, then

any center-connected star network L′ ∈ LCCS is also robust to overinvestment within group.

Moreover, if L 6∈ LCCS, then for a range of parameter specifications any center-connected

star network L′ ∈ LCCS is robust to overinvestment within group but L is not.

The proof of Proposition 8 is in Appendix C. In Proposition 6 we found that when all

agents are from the same group, incentives for overinvestment (within group) are minimized

by forming a (within-group) star. However, the incentives to form superfluous within-group

links are weakly greater when someone within the group holds an across-group link (see

equation 12). We can therefore think of the incentives for over-investment we found in

Proposition 6 as a lower bound on the minimal incentives we can hope to obtain once there

are across-group links. A key step in the proof of Proposition 8 shows that this lower bound

is obtained by all center-connected star networks.

Consider a center-connected star network L′. As the agent at the center of a within-group

star, agent k, has a link to all agents within the same group, we can focus on the incentives

of two non-center agents from the same group, i and j, to form a superfluous link. Consider

any subset of agents S ⊆ N such that i, j ∈ S. On the induced subnetwork L′(S) either lij is

superfluous or else k 6∈ S. This implies that no across-group links are present whenever the

additional link lij makes a positive marginal contribution. Hence considering different arrival

order, the average marginal contribution of such a link when it is added is the same on the

star network with no across-group links as for a center-connected star network: The lower

bound on within-group overinvestment incentives is obtained.

We now consider the within-group network structures that maximize the incentives for an

across-group link to be formed. We have already established that the marginal contribution

of a first bridging link to the total surplus is increasing in the sizes of the groups it connects.

41In the case of an essential across-group link that bridges two otherwise disconnected groups, the comparativestatics are unambiguous. In this case, the variance reduction is increasing in the sizes of the groups connected,as shown by inequality (4).


By the Myerson calculation, the agents with the strongest incentives to form such links are

then those who will be linked to the greatest number of other agents within their group when

they arrive. The result below formalizes this intuition.

Let A(Sk) be the set of possible arrival orders for the agents in Sk. For any arrival order

A ∈ A(S), let Ti(A) be the set of agents to whom i is path connected on L(S′), where S′ is

the set of agents (including i) that arrive weakly before i. Let T(m)i be a random variable,

taking values equal to the cardinality of Ti(A), where A is selected uniformly at random from

those arrival orders in which i is the m-th agent to arrive.

We will say that agent i ∈ Sk is more Myerson central (from now on, simply more central,

for brevity) within his group than agent j ∈ Sk if T(m)i first-order stochastically dominates

T(m)j for all m ∈ 1, 2, ..., |Sk|.42 In other words, considering all the arrival orders in which i

is the m-th agent to arrive, and all the arrival orders in which j is the m-th agent to arrive,

the size of i’s component at i’s arrival is larger than that of j’s at j’s arrival in the sense of

first-order stochastic dominance.43 This measure of centrality provides a partial ordering of

agents.

Lemma 9. Suppose agents in S0 form a network component, and all other agents in N form

another network component. Let i, i′ ∈ S0 and let j 6∈ S0. If i is more central within group

than i′, then i receives a higher payoff from forming lij than i′ receives from forming li′j:

MV (i;L ∪ lij)−MV (i;L) > MV (i′;L ∪ li′j)−MV (i′;L)

The proof is relegated to Appendix C. The key step in the proof pairs the arrival orders

of a more central agents with a less central agent, so that in each case the more central agent

is connected to weakly more people in the same group upon his arrival, and to the same set

of people from other groups. Such a pairing of arrival orders is possible from the definition

of centrality, and in particular the first-order stochastic dominance it requires.

Lemma 9 shows that more central agents have better incentives to form intergroup links.

We can then consider the problem of maximizing the incentives to form intergroup links by

choosing the within-group network structures (networks containing only within-group links).

We will say that the within-group network structures that achieve these maximum possible

incentives are most robust to underinvestment inefficiency across groups.

Proposition 10. If any efficient network L is robust to underinvestment across group, then

some center-connected star network L′ ∈ LCCS is also robust to underinvestment across

group. Moreover, if L 6∈ LCCS, then for a range of parameter specifications the center-

connected star network L′ ∈ LCCS is robust to underinvestment across group but L is not.

42We also use this notion of centrality to compare the within-group centrality of the same agent on twodifferent network structures. To avoid repetition we do not state the slightly different definition that wouldapply this situation.43An alternative and equivalent definition is that i is more central than j if there exists a bijection B : A(Sk)→A(Sk) such that |Ti(A)| ≥ |Tj(B(A))| and A(i) = A′(j), where A(i) is i’s position in the arrival order A andA′ = B(P ).


3 11

1

5

13 16

15 14 9

7

12

2

10

8

4

6

Figure 3. Center-connected within-group stars, in a context with two groups.

The proof of Proposition 10 is in Appendix C. Intuition can be gained from Lemma 9.

This Lemma shows that agents have better incentives to provide a bridging link across group

when they are more central within their own group. Thus to maximize the incentives of an

agent to provide an across-group link, we need to maximize the centrality of this agent within

group. This is achieved by any network that directly connects this agent to all others in the

same group. However, only one of these within-group network structures can be part of an

efficient network, and this is the star network, with the agent providing the across-group link

at the center.

Figure 3 shows a center-connected star network when there are groups. As long as it is effi-

cient for these groups to be connected, center-connected star networks and only the center-

connected star networks minimize the incentives for within-group overinvestment (by Propo-

sition 9) and minimize the inventives for across-group underinvestment (by Proposition 10).

The above results further reinforce the tension between efficiency and equality. The within-

group star not only minimizes the incentives for within-group overinvestment, it also mini-

mizes the incentives for across-group underinvestment. If an agent i provides an across-group

link, then of all the possible within-group network structures, the star with i at the center

maximizes i’s incremental payoff from establishing the link.44 More generally, the results

in this section also reinforce the main message of our paper: Once investments into social

capital are endogenized, new downsides to informal risk sharing agreements emerge.

6. Comparative Statics

We now turn our attention to taking our model to the data. The model is intentionally

stylized and intended to illustrate the key economic forces in the simplest possible way.

To take it to data we need to consider the general insights. The broad predictions of our

model thus far are that (1) there is endogenous centrality, (2) there is no underinvestment

within groups, (3) agents cannot be too far away from each other (in terms of the Myerson

distance), (4) agents that have across-caste links should be more Myerson central. However,

44Nevertheless, for some (but not all) parameter values, a within-group star will be more equitable when thecentral agent forms an across-group link than without it, because the across-group link generates positivespillovers to the whole group.


these predictions provide neither a clean nor powerful test of our theory. Predictions (2) and

(3) depend on an unobservable linking cost parameter and there are many alternative stories,

including ones not directly connected to risk sharing, that generate similar predictions. For

example, if individuals have heterogenous time budgets and made random links within and

across groups, predictions (1) and (4) are mechanically generated.

We therefore turn to more subtle predictions that rely on the comparative statics under

our model as we change parameters of the economic environment. We study four demanding

predictions from the theory with richer empirical content.

The first two predictions look at how network structure, described by Myerson distances

among agents, depends on income variability and correlation. The intuition is that in net-

works where the gains from risk sharing are higher (income is more variable/less highly

correlated), the Myerson distance between any two agents cannot be too high in a stable

network. Otherwise, a pair of individuals would be incentivized to form an additional link.

When the gains from risk sharing are lower, the Myerson distance can be larger.

The latter two predictions describe the composition of across-group bridging links. Individ-

uals with higher Myerson centrality have better incentives to form bridging links. However,

when income variability or the within- versus across-group income correlation is high, many

members of either group can find it worthwhile to form a cross-group bridging link. Thus,

we expect the average centrality within their own group of forming the cross-group link is

lower when the incentives to form such links are higher. We now formalize our predictions.

In the case of one group, Proposition 4 provides the key characterization of the set of

pairwise-stable networks. This characterization yields an exact expression for increased pay-

offs two agents would receive were they to form an additional link. For a risk-sharing network

to be stable, these benefits should be less than the cost of forming the link: for every i and

j that do not have a link,

(15) md(i, j, L) ≤ 2κw(1− ρw)λσ2

.

In our empirical setting, described below, we are interested in Indian village networks

where there are multiple groups, given by caste. Nevertheless, inequality (15) provides an

appropriate benchmark for within-group links. Recall that the left-hand side of the inequality

captures the probability that the link is not essential in the case of a random arrival order and

the right-hand side gives the value of the variance reduction obtained. The key complication

is that when a within-group link is essential for a subgraph but it connects two otherwise

separate components that contain people from multiple groups, then the value of the variance

reduction will depend on the composition of the people within the two components.

Consider the variance reduction obtained by combining any two components, with agents

in groups S0, . . . ,Sk and S0, . . . , Sk respectively. From equation 12, for some function f , the

certainty-equivalent value of this variance reduction is


[(1− ρw) + f(s0, . . . , sk, s0, . . . , sk)(ρw − ρa)]λσ2

2.

This is hard to compute in general. We approximate it by assuming that within each compo-

nent there are the same number of people from each group: si = α and si = β for i = 0, . . . , k.

Proposition 11. The certainty-equivalent value of variance reduction obtained by linking a

component with α people from each of groups 0, ..., k to a component with β people from each

of groups 0, ..., k is

(1− ρw)λσ2

2.

The proof is in Appendix C. Proposition 11 shows that the variance reduction obtained

by permitting two components containing agents from multiple groups to share risk is the

same as when all agents are from the same group, as long the proportion of people from each

group is the same in each component.

The above considerations lead to the following predictions:

P1. In villages with higher σ2, the average Myerson distance should be smaller.

P2. In villages with lower ρw , the average Myerson distance should be smaller.

While our first set of predictions looks at the relationship between the Myerson distance

and the environmental parameters, our second set of predictions looks at the composition of

the links. Our interest is in which agents provide the across-group links. Proposition 9 shows

more central agents have better incentives to provide an across-group link. The importance

of centrality will depend on the overall strength of the incentives to form across-caste links.

When income variance is high, or within-caste income correlation is high relative to across-

caste income correlation, the incentives to form an across-caste link will also be high, and so

network position will be less important; villagers in more varied locations will have sufficient

incentives to form across-caste links. More formally, from the variance reduction given in

equation 12 it is straightforward to show that for an across-caste bridging link lij :

∂∆ Var(L,L ∪ lij)∂σ2

> 0,∂∆ Var(L,L ∪ lij)

∂ρw> 0,

∂∆ Var(L,L ∪ lij)∂ρa

< 0.

This means that the incentives to form an across-caste link are increasing in σ2 and ρw−ρa,leading to the following predictions:

P3. In villages with higher σ2, the association between within-group centrality and the

formation of across-group links is lower.

P4. In villages with higher ρw − ρa, the association between within-group centrality and

the formation of across-group links is lower.

P1-P4 provide a complex set of predictions that we test empirically in the next section.

We can think about villages where households experience more variable income, e.g., due

to rainfall, or more correlated income and ask whether changes in income variability or

correlation correspond to changes in network structure.


7. Empirical Analysis

We now turn to data and see if there is support consistent with the above predictions.

We make use of a unique and detailed social network dataset from 75 villages in Karnataka,

India, which is particularly well-suited for our analysis as (i) it involves numerous independent

villages (essential for inference, though most network-based studies have just one or a handful

of villages), (ii) it includes complete network data across both financial and social connections

across almost all households in every village (network-based studies are notoriously subject

to measurement error), and (iii) caste is salient in these communities. A limitation of the

dataset is the lack of detailed financial and income data, but the data is unique in the sense

that it allows us to study many independent networks which is necessary for our analysis.

7.1. Setting and Data. The data we use were collected in 2011 by Banerjee, Chandrasekhar,

Duflo and Jackson (2013a, 2014a), by conducting surveys in 75 villages in Karnataka, India.

The villages span 5 districts and range from 2- to 3-hour drive from Bangalore. They are far

enough apart to be treated as independent systems (the median distance between them is 46

km, and a district has 1000–3000 villages). The survey included a village questionnaire, a

census of all households, demographic covariates (including caste and occupation), as well as

data on a number of amenities (e.g., roofing, latrine, or electricity access quality). A detailed

individual-level survey was administered to most adults in every village. The survey included

a networks module with twelve dimensions of relationships, including financial relationships,

social relationships, and advice relationships.45

Our analysis focuses on two types of networks: the financial graph, LF , and the social

graph, LS . The financial graphs represent risk-sharing connections, and the social graph

represents friendships/links used to socialize. We build “AND” networks, which say a link

exists if it exists on every dimension being considered (various types of financial connections

on the financial network, and various types of social connections on the social network).46

The advantage of doing this is that it generates a network structure that is more robust

to independent measurement error.47 The event that a frivolous link is coded decreases

exponentially if we require that it exist across multiple dimensions, which a priori helps in

detecting effects when we look differentially across network-type. In some of our empirical

analysis, we will explicitly consider how our predictions differentially play out in LF relative

to LS , as our theory speaks to the former.

Table 1 provides some summary statistics for our data. The average number of households

per village is 209 with a standard deviation of 80. The average degree of the financial graph

is 3.3, and of the social graph is 3.6. Further, we see that the clustering in the financial graph

45See Banerjee et al. (2014a) for more details. In total we have network data from 89.14% of the 16,476households based on interviews with 65% of all individuals of age 18 to 55.46We say lij ∈ LF if i goes to j to borrow money in times of need; j goes to i to borrow money in times ofneed; i goes to j to borrow material goods such as kerosene, rice, or oil in times of need; and j goes to i toborrow material goods in times of need.47We say lij ∈ LS if i goes to j’s house to socialize or vice versa.


is 0.19 (0.09), whereas for the social graph it is 0.11 (0.04). The fact that the clustering is

nearly twice as high for the financial graph is of note. This is consistent with the results of,

for instance, Jackson et al. (2012) that financial links may need to be supported/embedded

in cliques to sustain cooperation. Both the financial and social networks exhibit relatively

few cross-caste links. This is seen by looking at the ratio of the probability of having a

cross-caste link relative to the probability of having a within-caste link. A within-caste link

is three times as likely in the financial graph and less than twice as likely in the social graph.

Finally, 67% of households are high caste (GM or OBC).

Our analysis will focus on how the Myerson distance of the social network (for P1 and

P2) and Myerson centralities of those who link across caste (for P3 and P4) change with

parameters of the economic environment (σ2, ρw, and ρa). We develop an approximation

algorithm for our outcome variables – the Myerson distance and Myerson centrality – as

well as describe the construction of income variability (σ2) and the within- and across-group

income measures (ρw and ρa) in Appendix A.

7.2. Defining Groups and Boundaries of the Network. As our theory pertains to net-

works formed from multiple groups, here we make use of caste. Following Munshi and Rosen-

zweig (2006) and Banerjee et al. (2013), we partition our individuals into two broad caste

groups: scheduled caste/scheduled tribes (SC/ST) and general merit/otherwise backward

castes (GM/OBC). These are governmental designations used to condition the allocation of,

for instance, school seating by caste and reflect a core fissure in the social fabric.

It is worth noting that our model makes simplifying assumptions that do not perfectly fit

the data, and they can at most approximate it. One of these is that the linking costs between

any two individuals in the same group are the same, and the linking cost between any two

individuals from any two different groups are the same. These clearly do not hold exactly in

our setting, no matter how we define groups.

An extensive literature documents that in India, caste plays a significant role in informal

risk sharing. For instance, Morduch (1991, 1999, 2004) and Walker and Ryan (1990) show

that informal insurance functions rather well within a caste (though it still is imperfect),

while there is very limited insurance across caste. Morduch (2004) discusses this literature

at large, which primarily uses the methodology developed in Townsend (1994) to describe

the extent of risk sharing within and across caste groups in a village. More recent work by

Munshi and Rosenzweig (2014) as well as Mazzo and Saini (2012) has begun to turn attention

to risk-sharing at the level of subcastes, both within and across villages.

Our motivation for choosing caste (as opposed to for example sub-caste) as the defining

unit of group comes from multiple sources. First, an extensive sociological literature (see

for example, Srinivas (1962)) argues that it is particularly costly for people to form cross-

caste links. Second, anecdotal evidence suggests that our chosen caste group division is the

principle social fissure in these communities (e.g., evidenced by SC/ST “colonies” or sectors


of the village).48 Third, defining two large groups allows us to say more direct and precise

statements about the network structure within and across each group.49 Finally, we note that

this discussion only pertains to P3 and P4, where we are interested in how the within-group

centrality of a node making an across-group link varies with economic parameters. P1 and

P2 look at the network as a whole, so the only role caste boundaries play is in defining the

within- and across-group income correlation. The results of P1 and P2 are entirely unaffected

by whether we compute these quantities at the caste or subcaste level and we present results

where we compute the correlation at the sub-caste level.

It is well-understood in the literature that an important component of risk-sharing comes

from members of one’s subcaste, even across villages (see Munshi and Rosenzweig, 2014).

From the vantage point of this paper, we take the stance that establishing outside links

and outside risk-sharing arrangements takes a longer time than within village ones, because

of the lower frequency of interaction with those outside agents. Because of this we take

those arrangements given and incorporated into income realizations of agents in the village

by the time our agents start interacting. This is of course a simplifying assumption, but

we regard it as a reasonable approximation of reality. Restricting attention to risk-sharing

within a village is also consistent with what is done in the literature (see, e.g., Townsend

(1994), Kinnan (2011), Chiappori, Samphantharak, Schulhofer-Wohl and Townsend (2012),

and Ambrus et al. (2014), among others).

7.3. Empirical Strategy. Our predictions are about how network structure varies with σ2,

ρw, and ρa. The analysis is observational (not causal) and simply looks at the cross-sectional

variation of network structure with these parameters through ordinary least squares (OLS).

We take two approaches to support our empirical claims. First, by focusing on different

aspects of network structure (Myerson distance or the composition of cross-caste links), we

provide evidence in support of our predictions from very different moments of the data.

Second, we exploit multigraph data. Our theory is built for risk-sharing networks and

not for social connections. Since the theory is differentially more informative about the

structure of financial links as opposed to social links, the difference in patterns across link

types is informative. Furthermore, as unobserved endogeneity or homophily is likely to affect

several dimensions of the multigraph at once, looking at the difference in risk-sharing link

48We think focusing on caste in this aggregate way is a reasonable approximation to the setting described inour theory. For example, agents are about 3 times as likely to link within caste block than across caste block.They are also about 3 times as likely to link across subcaste, but within caste, as compared to across caste.On the other hand, they are twice as likely to have a within subcaste link as compared to a within caste,across subcaste link. It is also worth noting that if we partition all households into subcates, in the typicalvillage over 85% of nodes are members of at most 3 subcastes, along with a smattering of smaller subcastes.49Cutting the sample into very fine groups of can amplify measurement error. With a subset of only 5 or10 nodes, the notion of who is central and who isn’t is remarkably sensitive to survey measurement error.Unsurprisingly, the results for P3 and P4 become very noisy when we repeat the analysis at the sub-castelevel.


patterns versus social link patterns within a village allows us to address unobserved village-

level endogeneity that enters additively through a fixed effect. To take a simple example for

a confounder, consider P1. In villages where the weather is more variable, fewer days are

suitable for working, so people may spend more time socializing, thereby spuriously generating

shorter Myerson distances on average. Our approach eliminates this sort of confound.

Our empirical approach is more conservative than similar studies in the literature (e.g.,

Karlan et al. (2009), Ambrus et al. (2014), Kinnan and Townsend (2012)) in terms of statisti-

cal inference. First, these studies typically have very few networks (2, 1, and 16, respectively),

and therefore consider node or link-level regressions with standard errors generated at that

level. This effectively treats nodes or dyads as independent or loosely correlated, making in-

ference preclude, essentially, village-level shocks. Correlation in any factors outside the model

(e.g., incentives to form links for other reasons) as well as equilibrium selection at the vil-

lage level are all precluded from econometric analyses that don’t study the theory where the

entire graph is the unit of observation. We make no such assumptions on the independence

of nodes or dyads for valid statistical inference, and instead exploit the fact that we have 75

independent villages (recall that the median pairwise distance between them is over 46 km).

By focusing on village-level variation, we are allowing for arbitrary correlation within graphs.

In fact, in our most conservative specifications, we allow correlation at the subdistrict level.

Second, analyses in the previous literature usually do not have access to different types of

edges—the multigraph—and therefore cannot employ our difference-in-differences approach.

What we do, relative to this, is extremely conservative. By differencing across network types,

we are asking whether the patterns in the graph which match our theory are differentially at

play for the financial network relative to the social networks.

Third, typical models of multigraph link formation have a fixed-cost component. Incentives

driving the formation of risk-sharing links are likely to influence the structure of social or

information links through this channel. Observe that by looking at the financial graph relative

to the social graph, the variation coming from the fixed-cost component – which is consistent

with our theory – is not even being used in our analysis in support of our theory.

In sum, our analysis is observational, but makes use of variation in the economic environ-

ment as well as the type of network we focus on, to grapple with more correlated unobservables

(e.g., village level fixed unobservables) in a way that previous analysis could not.

7.4. Results.

7.4.1. Myerson distance as a function of income variability and correlation. We begin with

P1 and P2. We use a difference-in-differences approach and see whether the effects we are in-

terested in are coming differentially from the financial graph as opposed to the social graph.50

Figure 4 presents the raw data, differenced, graphically. Villages in areas corresponding to

more correlated within-caste income processes are associated with greater md(LF )−md(LS).

50We also run regressions without differencing out the social network and get similar results.


Similarly, an increase in income variability is associated with a differential decrease in the

measure of network density in the financial graph as compared to the social graph.

(a) md(LF )− md(LS) vs σ2 (b) md(LF )− md(LS) vs ρw

Figure 4. Myerson distance, differenced across network type, versus incomevariability and within-caste income correlation

Let v index village and t ∈ F, S index network type. We use the following regression:

(16) Yv,t = α+ βσv1t=F + γρw,v1t=F + µv + µt + δXv1t=F + εv,t,

where Yv,t = md(Ltv), µv is a possibly endogenous village fixed effect, µt is a network type

fixed effect, and Xv are village-level controls. P1 and P2 correspond to β < 0 and γ > 0, since

our theory pertains only to risk sharing networks. Thus, our effects should be differentially

more predictive for the financial network than for the social network and should remain true

when looking at relative effects. In all specifications here and throughout the paper, unless

otherwise noted we use a Wild clustered bootstrap to account for subdistrict-level clustering

in our inference (Cameron et al., 2008).51

Table 2 presents the results for two different measures of income correlation and two

approximations of the average Myerson distances between people. We find that a one standard

deviation increase in income variability differentially decreases md(L) by about 0.24 standard

deviations more in financial networks than in social networks (columns 1-4), irrespective of

the measure of income correlation used. We are unable to detect a statistically significant

association between rainfall variability and the outcomes of interest.

Thus, even when we look within villages, by allowing for village fixed effects and yet

allowing for subdistrict-level correlation in our error terms, we see that the financial graph

behaves in a manner more consistent with the theory than the social graph.

7.4.2. Association between within-caste centrality and cross-caste links. We now look at how

the composition of cross-caste links varies with these parameters. P3 says that in villages

51While these villages are essentially independent units, the median distance being 46 km apart, we take aconservative approach because geography is a determinant of rainfall variability and occupation. Our villagersare members of 12 subdistricts, and we therefore cluster our standard errors at that level. To deal withthe finite sample bias, we use a Wild cluster bootstrap procedure for the t-test statistic, using Rademacherweights, to generate p values for hypothesis testing as well as standard errors.


with higher variability we should see a greater association between within-caste centrality

and having an across-caste link. P4 says that a similar effect occurs when we look at within-

versus across-caste income correlation.

We present regressions of the form

(17) Yv,t = α+ βσv1t=F + γ (ρw,v − ρa,v) 1t=F + µv + µt + δXv1t=F + εv,t,

where Yv,t is the average within-caste centrality of nodes with a cross-caste link in network

t. The remaining terms are as before. P3 and P4 imply β < 0 and γ < 0.

Table 3 presents the results. Columns 1–3 use one approximation of average Myerson

distances, while columns 4–6 use another. The main result is that we find a one standard

deviation increase in the measure of the ρw − ρa, irrespective of which measure is used,

corresponds to about a 0.15 standard deviation decline in the average Myerson centrality

(within caste) of the individuals with across caste links. Again, however, when we look at

rainfall variability, we are unable to confirm or reject our theory, as estimates are very noisy.

Thus, our evidence is partially in support of our theory. We find a negative association

between income correlation and the centrality of those with a cross-caste link. However, we

are unable to confirm or reject our hypothesis when looking at the variability of income.

7.5. Discussion. The preceding analysis demonstrating the following:

(1) Networks have higher Myerson distances when within-group income correlation is

higher, though we detect no relationship with variability of rainfall.

(2) Networks exhibit more negative associations between within-group centrality and

across-group linking when within-group income correlation is higher, though we detect

no relationship with variability of rainfall.

(3) Results (1) and (2) are robust to a difference-in-differences approach. By differenc-

ing out the social graph, we can remove village-fixed endogenous factors and more

convincingly argue that the underlying force concerns risk sharing.

Having several distinct predictions is useful, as we are conducting an observational analysis.

We find that we cannot reject any of these predictions based on our data, and in fact, given

the data limitations we view our results as surprisingly supportive of the theory.

Our compositional story is new. More generally, P3–P4 are specific to either risk-sharing

relationships between castes or differences in income correlations between castes, and so

within the literature are unique to our model. These predictions are also more subtle. For

instance, higher within-group correlation leading to more sparse risk-sharing networks is not

a conclusion of other risk-sharing network formation stories.

We again emphasize that our findings are clearly observational (non-causal) and subject

to measurement error. However, given the unique dataset we used—network data for 75

independent villages, including multigraph data that allows us to difference out by link-

type to see if effects are driven by those consistent with our theory—this represents a first


opportunity to tackle the types of questions we address. The data requirements are immense,

and we are able to take a serious pass at looking at different types of cuts of our data—several

network measures, several parameters of the economic environment, and several different

results from our theory—to see if the data are consistent with our story. Additionally, we are

extremely conservative in conducting statistical inference, by relying only on independence

across villages (in fact, we are employing an even more conservative approach than that

by using a Wild cluster bootstrap at the subdistrict level) and not exploiting any of the

within-village observations (which are likely to be correlated).

8. Conclusion

We develop a relatively tractable model of network formation and surplus division in a

context of risk sharing that allows for heterogeneity in correlations between the incomes of

pairs of agents. Such correlations have a sizeable impact on the potential of informal risk

sharing to smooth incomes. We investigate the incentives for relationships that enable risk

sharing to be formed both within a group (caste or village) and across groups, giving access

to less correlated income streams. We find that overinvestment into social relations is likely

within a group, but there is potential underinvestment into more costly social connections that

bridge different groups. We also find a novel trade-off between equality and efficiency. Thus

we identify new downsides to informal risk sharing arrangements that can have important

policy implications.

Using a unique dataset of 75 Indian villages, we find empirical support for our model.

First, we find that within group income correlation and, to a lesser extent, income variabil-

ity are associated with the Myerson distance of the social network in a manner consistent

with our theory. Second, more centrally connected individuals are more likely to establish

across-group links when looking at networks with stronger within-group income correlation

relative to across-group income correlation. Third, each these relationships are differentially

stronger for the network of financial relationships than for the network of other types of social

relationships. The evidence, while descriptive, is congruent with our model.

Although we focus our analysis on risk sharing, our conclusions regarding network forma-

tion could apply in other social contexts too, as long as the economic benefits created by the

social network are distributed similarly to the way they are in our model—a question that

requires further empirical investigation. Within the context of risk sharing, a natural next

step would be to provide a dynamic extension of the analysis that allows for autocorrelation

between income realizations.


References

Ahuja, G. (2000): “Collaboration networks, structural holes, and innovation: A longitudinal

study,” Administrative science quarterly, 45, 425–455.

Algaba, E., JM Bilbao, J. Fernandez, N. Jimenez and J. Lopez (2007): “Algorithms for

computing the Myerson value by dividends,” Discrete Mathematics Research Progress, 1-13.

Ali, N. and D. Miller (2013): “Enforcing cooperation in networked societies,” mimeo UC San

Diego.

Ali, N. and D. Miller (2013): “Ostracism,” mimeo UC San Diego.

Ambrus, A., M. Mobius, and A. Szeidl (2014): “Consumption risk-sharing in social networks,”

American Economic Review, 104, 149-182.

Angelucci, M. and G. De Giorgi (2009): “Indirect effects of an aid program: How do cash

injections affect non-eligibles’ consumption?,” American Economic Review, 99, 486–508.

Arrow, K. (1965): “Aspects of the theory of risk-bearing (Yrjo Jahnsson Lectures),” Yrjo

Jahnssonin Saatio, Helsinki.

Atkinson, A. B. (1970): “On the Measurement of Inequality,” Journal of Economic Theory,

2, 244-263.

Attanasio, O., A. Barr, J. Cardenas, G. Genicot and C. Meghir (2012): “Risk pooling, risk

preferences, and social networks,” AEJ - Applied Economics, 4, 134-167.

Austen-Smith, D. and R. Fryer (2005): “The economic analysis of acting white,” Quarterly

Journal of Economics, 120, 551-583.

Bala, V. and S. Goyal (2000): “A noncooperative model of network formation,” Econometrica,

68, 1181-1229.

Banerjee, A., A. Chandrasekhar, E. Duflo and M. Jackson (2013a): “The diffusion of micro-

finance,” Science, 341, No. 6144 1236498

Banerjee, A., E. Duflo, M. Ghatak, and J. Lafortune (2013b): “Marry for what? Caste and

mate selection in modern India.” American Economic Journal: Microeconomics, 5(2): 33-72.

Banerjee, A., A. Chandrasekhar, E. Duflo and M. Jackson (2014a): “Gossip: Identifying

central individuals in a social network,” MIT and Stanford working paper.

Banerjee, A., A. Chandrasekhar, E. Duflo and M. Jackson (2014b): “Exposure to formal

loans: Changes in social networks,” MIT and Stanford working paper.

Banerjee, A., E. Breza, E. Duflo and C. Kinnan (2014c): “Does microfinance foster business

growth? The importance of entrepreneurial heterogeneity,” mimeo MIT.

Bennett, E. (1988): “Consistent Bargaining Conjectures in Marriage and Matching,” Journal

of Economic Theory, 45, 392-407.

Bergstrom, T. and H. Varian (1985): “When do markets have transferable utility?,” Journal


Billand, P., Bravard, C. and S. Sarangi (2012): “Mutual insurance networks and unequal

resource sharing in communities,” DIW Berlin Discussion Paper.


Binzel, C., E. Field and R. Pande (2014): “Does the arrival of a formal financial institution

alter informal sharing arrangements? Evidence from India,” mimeo Harvard University.

Bloch, F., G. Genicot, and D. Ray (2008): “Informal insurance in social networks,”Journal


Borch, K. (1962): “Equilibrium in a reinsurance market,” Econometrica, 30, 424-444.

Bramoulle, Y. and R. Kranton (2007a): “Risk-sharing networks,”Journal of Economic Be-

havior and Organization, 64, 275-294.

Bramoulle, Y. and R. Kranton (2007b): “Risk sharing across communities,” American Eco-

nomic Review Papers & Proceedings, 97, 70.

Burt, Y. (1992): “Structural holes,” Academic Press, New York.

Calvo-Armengol, A. (2001): “Bargaining power in communication networks,” Mathematical

Social Sciences, 41, 69-87.

Calvo-Armengol, A. (2002): “On bargaining partner selection when communication is re-

stricted,” International Journal of Game Theory, 30, 503-515.

Calvo-Armengol, A. and R. Ilkilic (2009): “Pairwise-stability and Nash equilibria in network

formation,” International Journal of Game Theory, 38, 51-79.

Cameron, A., J. Gelbach and D. Miller (2008): “Bootstrap-based improvements for inference

with clustered errors, ” The Review of Economics and Statistics, 90(3), 414-427.

A. Chandrasekhar and R. Lewis (2014): “Econometrics of sampled networks,” mimeo Stan-

ford University.

Cole, H., and N. Kocherlakota (2001): “Efficient allocations with hidden income and hidden

storage,” The Review of Economic Studies 68, 523-542.

Coate, S. and M. Ravaillon (1993): “Reciprocity without commitment: Characterization and

performance of informal insurance arrangements,” Journal of Development Economics, 40,

1–24.

Corominas-Bosch, M. (2004): “Bargaining in a network of buyers and sellers,” Journal of

Economic Theory, 115, 35-77.

Elliott, M. (2014): “Inefficiencies in networked markets,” AEJ Microeconomics, forthcoming.

Fafchamps, M. (1992): “Solidarity networks in preindustrial societies: Rational peasants with

a moral economy,” Economic Development and Cultural Change, 41, 147-74.

Fafchamps, M. and F. Gubert (2007): “The formation of risk sharing networks,” Journal of

Development Economics, 83, 326-350.

Feigenberg, B., E. Field and R. Pande (2013): “The economic returns to social interaction:

Experimental evidence from micro finance,” Review of Economic Studies, 80, 1459-1483.

Fontenay, C. and J. Gans (2013): “Bilateral bargaining with externalities,” Journal of Indus-

trial Economics, forthcoming.

Glaeser, E., D. Laibson, and B. Sacerdote (2002): “An economic approach to social capital,”

The Economic Journal, 112, F437-F458.


Goyal, S. and F. Vega-Redondo (2007): “Structural holes in social networks ,” Journal of

Economic Theory, 137, 460–492.

Hojman, D. and A. Szeidl (2008): “Core and periphery in networks,” Journal of Economic

Theory, 139, 295-309.

Jackson, M. (2003): “The stability and efficiency of economic and social networks,” in Ad-

vances in Economic Design, edited by Murat Sertel and Semih Koray, Springer-Verlag, Hei-

delberg.

Jackson, M. (2005): “Allocation rules for network games,” Games and Economic Behavior,

51, 128-154.

Jackson, M. (2010): “Social and economic networks,” Princeton University Press.

Jackson, M., T. Rodriguez-Barraquer and X. Tan (2012): “Social capital and social quilts:

Network patterns of favor exchange,” The American Economic Review 102.5 (2012): 1857-

1897.

Jackson, M. and A. Wolinsky (1996): “A strategic model of social and economic networks,”

Journal of Economic Theory, 71, 44-74.

Karlan, D., M. Mobius, T. Rosenblat and A. Szeidl (2009): “Trust and social collateral,”

Quarterly Journal of Economics, 124, 1307-1361.

Kets, W., G. Iyengar, R. Sethi and S. Bowles (2011): “Inequality and network structure,”

Games and Economic Behavior, 73, 215-226.

Kinnan, C. (2011): “Distinguishing barriers to insurance in Thai villages,” mimeo North-

western University.

Kinnan, C. and R. Townsend (2012): “Kinship and financial networks, formal financial access,

and risk reduction,” The American Economic Review, 102, 289-293.

Kranton, R. and D. Minehart (2001): “A theory of buyer-seller networks,” American Eco-

nomic Review, 91, 485-508.

Ligon, E., , J. Thomas, and T. Worrall (2002): “Informal insurance arrangements with limited

commitment: Theory and evidence from village economies,” Review of Economic Studies, 69,

209–244.

Manea, M. (2011): “Bargaining in stationary networks,” American Economic Review, 101,

2042-80.

Matsuura, K. and C.J. Willmott (2012): “Terrestrial precipitation: 1900–2008 gridded monthly

time series,”Center for Climatic Research Department of Geography Center for Climatic Re-

search, University of Delaware.

Mazzocco, M. and S. Saini (2012): “Testing efficient risk sharing with heterogeneous risk

preferences,”The American Economic Review, 102.1, 428-468.

Mehra, A., Kilduff, M. and D. Bass (2001): “ The social networks of high and low self-

monitors: implications for workplace performance,”Admin. Sci. Quart. 46 (2001) 121–146.

M. Morduch (1991): “Consumption smoothing across space: Tests for village-level responses

to risk.” Harvard University.


M. Morduch (1999): “Between the state and the market: Can informal insurance patch the

safety net?” The World Bank Research Observer 14.2, 187–207.

Munshi, K. and M. Rosenzweig (2006): “Traditional institutions meet the modern world:

Caste, gender, and schooling choice in a globalizing economy,” The American Economic

Review, 1225-1252.

Munshi, K. and M. Rosenzweig (2009): “Why is mobility in India so low? Social insurance,

inequality, and growth,”National Bureau of Economic Research, No. w14850.

Myerson, R. (1977): “Graphs and cooperation in games,” Mathematics of Operations Re-

search, 2, 225-229.

Myerson, R. (1980): “Conference structures and fair allocation rules,” International Journal

of Game Theory, 9.3, 169-182.

Myerson, R. (1991): “Game theory: Analysis of conflict, Harvard University Press, Cam-

bridge.

Navarro, N. and A. Perea, (2013): “A simple bargaining procedure for the Myerson value,”

The BE Journal of Theoretical Economics, 13, 1-20.

Pin, P. (2011): “Eight degrees of separation,” Research in Economics, 65, 259-270.

Podolny, J., and J. Baron, (1997): “Resources and relationships: Social networks and mobility

in the workplace,” American Sociological Review, 673-693.

Rosenzweig, M. (1988): “Risk, implicit contracts and the family in rural areas of low-income

countries,”Economic Journal, 98, 1148-70.

Slikker, M. (2007): “Bidding for surplus in network allocation problems,” Journal of Economic

Theory, 137, 493-511.

Srinivas, M. (1962): “Caste in modern India and other essays,” Bombay, India: Asia Pub-

lishing House.

Stole, L. and J. Zwiebel (1996): “Intra-firm bargaining under non-binding contracts,” Review

of Economic Studies, 63, 375-410.

Townsend, R. (1994): “Risk and insurance in village India,” Econometrica, 62, 539-591.

Udry, C. (1994): “Risk and insurance in a rural credit market: An empirical investigation in

Northern Nigeria,” Review of Economic Studies, 61, 495-526.

van Lint, J. H., and R. Wilson (2001): “A course in combinatorics,” Cambridge university

press.

Walker, T. and J. Ryan (1990): “Village and household economics in India’s semi-arid trop-

ics,” Johns Hopkins University Press, Baltimore.

Wilson, R. (1968): “The theory of syndicates,” Econometrica, 36, 119-132.


Table 1. Summary Statistics

Mean SDHouseholds per village 209.27 80.03Average degree (financial network) 3.28 0.91Average clustering (financial network) 0.19 0.05Probability of cross-caste link/ Probability of within-caste link (financial network) 0.35 0.30Average degree (social network) 3.55 1.15Average clustering (social network) 0.11 0.04Probability of cross-caste link/ Probability of within-caste link (social network) 0.62 0.49Fraction high caste (GM/OBC) 0.67 0.15Notes: For each variable we present the mean and the cross-village standarddeviation. For the Myerson distances we use the approximation algorithm developedin the paper.


Table 2. Myerson distance vs. income variability and correlation

Measure I Measure II Measure I Measure II(1) (2) (3) (4)

Income Correlation x 1Financial network 0.242 0.177 0.225 0.289(0.104) (0.115) (0.084) (0.115)[0.012] [0.188] [0.010] [0.016]

Income Variability x 1Financial network -0.070 -0.057 -0.151 -0.131(0.107) (0.115) (0.178) (0.171)[0.564] [0.712] [0.446] [0.502]

R-squared 0.8353 0.8297 0.7362 0.7452Notes: Outcome variable is the average Myerson distance of the graph of a given type (financial or social), computed using the specifiedalgorithm (incoming or outgoing link deletion). Outcome variable, income correlation (either Measure I or Measure II), and income variabilityall scaled by their standard deviations. With two observations per village, one for each network type, all regressions include village fixed effects(and therefore subsume district and subdistrict fixed effects as well). Specifications include control for caste composition: p(1-p), where p isshare in high caste, interacted with network type. Wild clustered bootstrap standard errors are presented in (.), using Rademacher weights. Thecluster is at the subdistrict level, of which there are 12. 1000 samples are used per bootstrap. p-values from the Wild clustered bootstrap t arepresented in [.].

Panel A: Incoming Link Deletion Algorithm Panel B: Outgoing Link Deletion Algorithm


Table 3. Association between average within-caste centrality of nodes withcross-caste links and measures of rainfall variability and within-caste incomecorrelation


Income Correlation x 1Financial network -0.191 -0.228 -0.121 -0.263(0.088) (0.150) (0.095) (0.118)[0.030] [0.12] [0.200] [0.026]

Income Variability x 1Financial network 0.056 0.060 0.109 0.104(0.162) (0.166) (0.188) (0.177)[0.727] [0.717] [0.564] [0.556]

R-squared 0.7956 0.7954 0.765 0.772

Incoming Link Deletion Algorithm Outgoing Link Deletion Algorithm

Notes: Outcome variable is the average within-caste average Myerson centrality, computed using our approximation algorithm (incoming linkdeletion or outgoing link deletion), of those with a cross-caste link. Income correlation given by Measure I or Measure II as indicated.Outcome variables and regressors are scaled by their standard deviations. The outcome variable is computed using the financial or socialnetwork (two observations per village for the difference-in-difference). All specifications use village fixed effects, and both panels includecontrols for caste composition and size (interacted with network type). Wild clustered bootstrap standard errors are presented in (.), usingRademacher weights. The cluster is at the subdistrict level, of which there are 12. 1000 samples are used per bootstrap. p-values from the Wildclustered bootstrap t are presented in [.].


Appendix A. Variable Construction

A.1. Approximating the Myerson distance and centrality. We need to compute the

Myerson distance of every pair in every village and the Myerson centrality for all nodes.

Unfortunately, this is computationally infeasible for the sample sizes of our data (see Algaba

et al. (2007)), presenting a new challenge. Thus, we develop an approximation, described

below.

Let md(L) be the matrix of Myerson distances and define q(L) := 1/2−md(L). So q(L)

is a matrix with the ijth entry capturing the probability that, upon his arrival agent i will not

be connected to agent j. It is difficult to directly characterize md(L) (or equivalently, q(L))

as each village typically consists of around 230 households and the number of candidate paths

between each i and j is exponential in the size of the network. Correctly accounting for paths

that share nodes is computationally very intensive (see Proposition 4), and it has to be done

for all pairs of agents without a direct connection.52 Instead, we develop a computationally

feasible approximation of md(L), which is exact for trees.

To approximate q, we use the following idea. The algorithm works by starting with a node,

moving to its neighbors, then move to its neighbors’ neighbors, and so on, never returning

to a previously used node along a given path. This helps us to avoid counting non-minimal

paths. All the while, we keep track of how many ways we have moved from the original node

to any given node. We denote our approximation of q by q.

1

2 3

4 5 6 7

8 9 10

11

(a) Tree

1 2

3

4 5

6

(b) Circle

Figure 5. The nodes i,j for which we are computing md(i, j, L) have purplestripes. The tree contains a single minimal path (solid orange nodes), whereasthe circle contains two paths (solid orange nodes and chequered blue nodes).

The inclusion–exclusion principle weights paths that are longer less and a path that shares

many nodes with another less. With this in mind, we choose the following two approximation

strategies. Let the shortest path between two nodes be of length l. We first count the paths

of length l and length l + 1. We then count paths of length l + 2.53 If there are fewer than

k such paths, we use them all. Otherwise, we consider only the k shortest and in practice

52Further, due to presumed measurement error (see Banerjee et al. (2013)), there are likely to be missingpaths. In fact, the data have occasional disconnected components, and so measures that are precisely basedon exact paths or even maximal path lengths are likely to be problematic (Chandrasekhar and Lewis (2014)).53Counting more paths greatly (exponentially) increases the running time of our algorithm.


we set k = 4.54 Discarding longer paths in this way biases downwards our approximation

of q. As we cannot keep track of exactly which nodes feature in each path, we also have

to make an assumption about the overlap of nodes in order to apply the inclusion–exclusion

principle to these paths. Each path must share the same first and last node. We perform

the inclusion–exclusion principle assuming that only these nodes are shared (see Section 4).

Assuming no other nodes are shared introduces a second bias, but this time upwards in our

approximation.

To explain these concepts, we provide some illustrations. Figure 5 presents two examples:

a tree and a circle. The tree has a single minimal path between nodes 1 and 8, whereas the

circle has two minimal paths between nodes 1 and 4. Figure 6 shows how links are removed

for the case of a tree. Once a node has been reached, links back into that node are deleted

before the nodes neighbors are “infected.” This ensures only minimal paths are included in

the calculation.

A

B

(a)

A

B

(b)

A

B

(c)

A

B

(d)

Figure 6. As the algorithm progresses, directed links into nodes that arereached are deleted. This ensures that only minimal paths are included. Inthis case, as in all tree networks, there is a unique minimal path from A to B.

In the case of the circle shown in Figure 5, our algorithm is also exact for paths between 1

and 4. There are two minimal paths (which in this case are both shortest paths too), and we

find both in the initial run of our algorithm. Following the inclusion–exclusion principle, we

add 1/4 to 1/4 and subtract 1/6. In this case our assumption that the two paths share only

two nodes is accurate. We are also exact for paths between 1 and 3, but in this case there is

a path of length l+ 2. To find this path, we look for paths of length l from 1 to nodes other

than 3. In this case there is one such path to node 5. We then look for paths from 5 to 3

that pass through one other node. There is one such path and so the calculation we perform

is: 1/3 + 1/5− 1/6. While we are accurate for all pairs of nodes in the circle shown, in larger

circles we will miss the longer paths.

54We need a fixed (small) truncation. Otherwise both the memory requirements and the run-time of thealgorithm grow exponentially. Results are not sensitive to the truncation point.


The following algorithm finds the length of the shortest path between two nodes, how

many paths of that length there are and how many paths there are that are one longer. From

this information, we also find paths of length l + 2.55

Algorithm 12 (Incoming Link Deletion). Let ei be the ith basis vector. This will represent

the root (starting) node. Initialize q = zeros(n, n), a matrix of zeros. Initialize zt,i =

zeros(n, 1) and xt,i = zeros(n, 1) to be n-vectors of zeros, indexed by i = 1, ..., n and t =

1, ..., T . Repeat steps 1–4 for each of (e1, ..., en).

(1) Period 1: There is no identification or updating steps.

(a) Percolation: x1,i = Aei.

(Identifies who is connected to the root node)

(2) Period 2, given (x1,i,A):

(a) Identification: z2,i = ei.

(b) Update graph:56 A2 = zeros(n, n), A2(¬z2,i, :) = A(¬z2,i, :).

(Deletes links into the root node)

(c) Percolation: x2,i = A2x1,i.

(Records number of paths from root node to other nodes passing through one

other)

(3) Period t, given (xt−1,i,At−1):

(a) Identification: zt,i = 1∑t

s=3 xs−2,i > 0

.

(Identifies nodes already visited)

(b) Update graph: At = zeros(n, n), At(¬zt,i, :) = At−1(¬zt,i, :).(Deletes links into all nodes that have already been visited)

(c) Percolation: xt,i = Atxt−1,i.

By construction xt,ij , the jth entry of xt,i, records paths from i to j that pass through t

nodes. If t′ is the lowest t with a positive entry in this matrix, then the shortest path from

i to j passes through t′ nodes. In this case, xt′,ij tells us how many such paths there are and

xt′+1,ij tells us how many paths there are that pass through one more node. However, by

construction xt′+k,ij = 0 for all k > 1 and longer paths are not recorded. This is because the

incoming links to node j will have been deleted by this step of the algorithm. Deletion of

incoming links helps prevents non minimal paths from being recorded. Using this information

for all seed nodes, the number of paths of length t′ + 2 between i and j are also found as

described above. The inclusion-exclusion principle is then applied to this combined set of

paths, assuming each path shares only the first and last nodes, to calculate q(L).

Proposition 13. Let L be a tree. Then q(L) = q(L).

55For paths from i to j, this is done by looking at paths of length l to agents other than j, and then lookingat paths from these agents to j.56Let A(:, v) denote (A(1, j), ..., A(n, j)).


See the Appendix C for the proof. To operationalize q(L) in our regression analysis,

we need a village-level measure of Myerson distances. We use q(L) :=∑

i<j qij/(n2

)which

measures an appropriately weighted density of the network. Finally, to approximate Myerson

centrality we use∑

j qij , as people are central when they are likely to be connected to others.

Thus, their qi terms are high (equivalently, they have a low Myerson distance to others).

A limitation of the Incoming Link Deletion algorithm is that longer paths are excluded. To

address this, we construct an alternative algorithm. This Outgoing Link Deletion algorithm is

identical to the one described, except that it deletes outgoing links instead of incoming links.

The Outgoing Link Deletion algorithm finds longer paths, and does an especially good job

of picking up longer paths that share few nodes with other paths. However, it also includes

additional short non-minimal paths and is not exact for tree networks. As longer paths

are found, we directly use the output of the algorithm without constructing any additional

longer paths. Nevertheless, for the set of paths we find, it is computationally infeasible

to compute the Myerson distances using the inclusion-exclusion principle. Censoring these

paths would defeat the point of the Outgoing Link Deletion algorithm. Instead, we use

an approximation of the inclusion–exclusion principle which makes the computation much

simpler. This approximation treats every path as completely independent, assuming that no

nodes are shared (even though we know at two must be). For example, if we find 3 paths

from i to j that pass through l nodes, l′ nodes and l′′ nodes respectively, our approximation

of qij will be 1/l + 1/l′ + 1/l′′.

A.2. Construction of income variability and caste–income correlation. Our analysis

requires estimates of σ2 and (ρw − ρa). However, due to data limitations—despite extremely

detailed multigraph network data across many villages, we lack financial records—we need

to construct measures of these quantities.

For income variability, we merge National Oceanic and Atmospheric Administration (NOAA)

data with our network data. Matsuura and Willmott (2012) construct a gridded monthly

time series of terrestrial precipitation from 1900 to 2010. We match this to our villages using

our GPS data. Once month fixed-effects are removed, the crucial variable is the standard

deviation of rainfall by village.

Measuring income correlation is difficult. Ideally, we would have time-series data on in-

comes of all households, as well as plausible instruments allowing us to calculate the exogenous

variation in income correlation both within and across castes for each village. While we do

not have access to income data, we do have detailed data on occupation.57 Thus, we make

57We recognize that occupations can have a choice component. Nevertheless, we proceed with these measuresfor three reasons. First, in rural villages the primary household occupation (agriculture or sericulture) is oftenpassed on through generations. Second, it is possible to show that under a natural model with endogenousselection into occupation, the within- versus across-group income correlations are captured by our occupationalchoice measures, whereas the choice of occupation does not generate spurious correlations with the networkin a manner consistent with P1–P4. Finally, this is the best possible approximation, given the severe incomedata limitations, as a necessary component of our analysis is the network data.


use of the relative within-caste to across-caste occupation correlation. The main idea is that

shocks to individuals will be more highly correlated when they have the same occupation.

We take two approaches to computing caste–income correlation.

One approach is to look at the correlation of being in the high-caste group with holding

a given occupation, for all occupations in our survey. We then take the weighted average of

these correlations, where the weight is the share of agents in the occupation. Thus, for a

given village we consider

(18) ρw − ρa :=

K∑k=1

corr (Caste,Occupationk) Pr (Occupationk) ,

where Caste is a 1 × n vector of sub-caste dummies and Occupationk is an n × 1 vector of

dummies for a household having a member in occupation k. This constructs a score which

is 0 if there is no correlation between sub-caste group and occupation, and 1 if caste group

perfectly predicts occupational choices.58

Another approach involves making simplifying assumptions about the structure of the

income process. If the income of an individual in a given caste and occupation can be thought

of as depending on a (caste–occupation)-specific mean, occupation-specific idiosyncratic iid

shocks, and individual-specific idiosyncratic iid shocks, where the occupation shocks all have

the same variance across occupations, then as shown in Appendix D, we can write

(19) ρw − ρa :=∑g

φgPr(oi,g = oj,g)− Pr(oi,A = oj,B),

where oi,g denotes the occupation of i in caste g ∈ A,B and φg is the population share

of caste g. In sum, both measures are intuitive, but imperfect, proxies for the regressors we

need in our analysis. Therefore we take a rough-and-ready approach, utilizing both in our

analysis:

(1) ρw − ρaI

=∑

g φgPr(oi,g = oj,g)− Pr(oi,A = oj,B).

(2) ρw − ρaII

=∑K

k=1 corr (Caste,Occupationk) Pr (Occupationk) .

58Unsurprisingly, the analogous variable constructed at the GM/OBC level has a correlation of 0.94 with theone constructed using caste.


Supplementary Appendix: For Online Publication Only

Appendix B. Over and Under Investment Examples

In this Appendix we provide an example of over-investment within group in the unique

stable network and a related example of underinvestment across group in the unique stable

network.

We begin by assuming there is one group with s members connected by a network L.

Equation 10 implies that Myerson distance of two agents i, j such that lij 6∈ L is greater than

1/2, while the Myerson distanace between i and j if they form the link lij would be 1/2.

Thus i and j’s gross payoff strictly increases if the link lij . So, for κw sufficiently close to

0, in all stable networks for any pair of agents i, j the link lij must be formed; The unique

stable network is the complete network and there is overinvestment.

Suppose now there two groups, g, g′ both with s members and keep the same parameter

values from the previous example. By equation 12 the incentives for form within group links

are weakly increased by any across group links. Thus in all stable networks the network

structures within-group must be complete networks; All possible within-group links must be

formed. Suppose these are the only links formed so that no across-group links are formed.

Denote this network L. From equation 14 the change in total variance achieved by connecting

an agent i from group g to an agent j from group g′ is strictly increasing in the size of

both groups s. Considering the Myerson value calculation (equation 5), this means that the

marginal contribution of the link lij to total surplus (the certainty equivalent value of the

variance reduction) is strictly greater on L∪lij than it is on any strict subgraph, including

all those formed when the later of i and j arrives in the Myerson calculation. This implies

that (MV (i;L ∪ lij)−MV (i;L)) + (MV (j;L ∪ lij)−MV (j;L)) < TS(L ∪ lij)− TS(L) for

all lij : i ∈ Sg, j ∈ Sg′ . So, setting κa such that

MV (i;L ∪ lij)−MV (i;L) +MV (j;L ∪ lij)−MV (j;L) < 2κa < TS(L ∪ lij)− TS(L),

the network L is the unique stable network and there is underinvestment (in across-group

links) in all stable networks.

Appendix C. Omitted Proofs

Proof of Proposition 1. To prove the first statement, consider villagers’ certainty-equivalent

consumption. Let K be some constant, and consider the certain transfer K ′ (made in all

states of the world) that i requires to compensate him for keeping a stochastic consumption

stream ci instead of another stochastic consumption stream c′i:


E[v(ci + K)] = E[v(c′i + K −K ′)]

− 1

λe−λKE[e−λci ] = − 1

λe−λKeλK

′E[e−λc

′i ]

eλK′

=E[e−λci ]

E[e−λc′i ]

K ′ =1

λ

(ln(E[e−λci ]

)− ln

(E[e−λc

′i ]))

(20)

This shows that the amount K ′ needed to compensate i for taking the stochastic con-

sumption stream c′i instead of ci is independent of K. As a villager’s certainty-equivalent

consumption for a lottery is independent of his consumption level, certainty-equivalent units

can be transferred among the villagers without affecting their risk preferences, and expected

utility is transferable.

Next, we characterize the set of Pareto efficient risk sharing agreements. Borch (1962) and

Wilson (1968) showed that a necessary and sufficient condition for a risk-sharing arrangement

between i and j to be Pareto efficient is that in all states of the world ω ∈ Ω,

(21)

(∂vi(ci(ω))

∂ci(ω)

)/(∂vj(cj(ω))

∂cj(ω)

)= αij

where αij is a constant. Substituting in the CARA utility functions, this implies that

e−λci(ω)

e−λcj(ω)= αij

ci(ω)− cj(ω) = − ln(αij)

λ

E[ci(ω)]−E[cj(ω)] = − ln(αij)

λci(ω)− cj(ω) = E[ci(ω)]−E[cj(ω)](22)

Letting i and j be neighbors such that j ∈ N(i), equation 22 means that when i and

j reach any Pareto-efficient risk-sharing arrangement their consumptions will differ by the

same constant in all states of the world. Moreover, by induction the same must be true for

all pairs of path-connected villagers.

Consider now the problem of splitting the incomes of a set of villagers S in each state of

the world to minimize the sum of their consumption variances:

(23) minc

∑i∈S

Var(ci) subject to∑i∈S

yi(ω) =∑i∈S

ci(ω) for all ω.

Note that,

(24)∑i∈S

Var(ci) =∑i∈S

∑ω∈Ω

p(ω)(ci(ω)−E[ci])2,


where p(ω) is the probability of state ω. As the sum of variances is convex in consumptions

and the constraint set is linear, the maximization is a convex program. The first-order

conditions of the Lagrangian are that for each i ∈ S and each ω ∈ Ω, 2(ci(ω)−E[ci]) = γ(ω),

where γ(ω) is the Lagrange multiplier for state ω. Thus, ci(ω)− cj(ω) = E[ci(ω)]−E[cj(ω)]

for all i, j ∈ S. This is exactly the same condition as the necessary and sufficient condition

for an ex-ante Pareto efficiency. Hence, a risk-sharing agreement is Pareto efficient if and

only if the sum of the consumption variances for all path-connected villagers is minimized.

Using the necessary and sufficient condition for efficient risk sharing, we obtain

(25)∑k∈S

yk(ω) =∑k∈S

ck(ω) = |S|ci(ω)−∑k∈S

(E[ci(ω)]−E[ck(ω)]) ,

which implies that

(26) ci(ω) =1

|S|∑k∈S

yk(ω) +1

|S|∑k∈S

(E[ci(ω)]−E[cj(ω)]) =1

|S|∑k∈S

yk(ω) + τi,

where τi = E[ci(ω)]−E[∑

k∈S yk(ω)].

Proof of Lemma 4. Agent i’s net benefit from forming link lij is (MVi(L)−MVi(L \ lij)−κw). We need to show that

(27) MVi(L)−MVi(L \ lij) = MVj(L)−MVj(L \ lij) = md(i, j, L)V.

Some additional notation will be helpful. Suppose agents arrive in a random order, with a

uniform distribution on all possible arrival orders. The random variable Si ⊆ N identifies the

set of agents, including i, who arrive weakly before i. For each arrival order, we then have an

associate network LL(Si) that describes the network formed upon i’s arrival (the subnetwork

of L induced by agents Si). Let q(i, j, L) be the probability that i and j are path connected

on network LL(Si).

The certainty-equivalent value of the reduction in variance due to a link lij in a graph

LL(Si) is V if the link is essential and 0 otherwise. The change in i’s Myerson value, MVi(L)−MVi(L \ lij), is then (q(i, j, L)− q(i, j, L \ lij))V . However, q(i, j, L) = 1/2. To see this,

note that lij ∈ L and therefore in every order of arrival in which i arrives after j (which

happens with probability 1/2), i and j are path connected on the network LL(Si), while i

and j are never path connected on LL(Si) when j arrives after i.

Probability q(i, j, L \ lij) can be computed by the inclusion-exclusion principle, using

the fact that the probability of a path connecting i and j existing on network LL\lij(Si) is

equal to the probability that for some minimal path connecting i and j on L\lij all agents

on the path are present in Si. Thus


(28) q(i, j, L \ lij) =

|P(i,j,L\lij)|∑k=1

(−1)k+1

∑1≤i1<···<ik≤|P(i,j,L)|

(1

|Pi1 ∪ · · · ∪ Pik|

) .

We therefore have that

(29) MVi(L)−MVi(L \ lij) = (1/2− q(i, j, L \ lij))V = md(i, j, L)V,

where the last equality follows from the definition of Myerson distance.

Proof of Proposition 5. For there to be underinvestment in a pairwise-stable network L, there

must exist a link lij 6∈ L for which TS(L∪ lij)− TS(L) > 2κw. This can only happen if lij is

essential on L∪ lij as otherwise TS(L∪ lij)−TS(L) = 0. Thus TS(L∪ lij)−TS(L) = V and

so V > 2κw. As lij 6∈ L and L is pairwise stable, Lemma 4 implies that md(i, j, L) ≤ κw/V .

However, as lij is essential on L ∪ lij , md(i, j, L) = 1/2. Substituting this into the condition

from Lemma 4 we get V ≤ 2κw, leading to a contradiction.

For the second part of the proposition, let L be a pairwise-stable network and let lij ∈ L be

an essential link on L in which there is overinvestment. Thus TS(L)− TS(L \ lij) = V <

2κw. Since lij is essential, md(i, j, L \ lij) = 1/2. But Lemma 4 implies that md(i, j, L \lij) ≥ κw/V . We therefore have that V ≥ 2κw, leading to a contradiction.

Proof of Proposition 6.

Part (i): By remark 3 and under our regularity condition, all efficient networks are tree

networks. By definition, in all tree networks any pair of agents i and j have a unique minimal

path between them. Thus, for a tree network L with diameter d(L), there exist agents i and

j with a unique minimal path between them of length d(L) and all other pairs of agents have

a weakly shorter minimal path between them. Thus by equation 10:

(30) md(i, j, L) =1

2− 1

d(L)≥ md(k, k′, L) for all k, k′ ∈ N.

By Proposition 5 there is no underinvestment in any stable network. Lemma 4 therefore

implies that the efficient network L is stable if and only if md(k, k′, L) ≤ κw/V for all k, k′

such that lij 6∈ L. As md(i, j, L) ≥ md(k, k′, L) and md(i, j, L) = 1/d(L) (see equation 30),

this condition simplifies and the efficient network L is stable if and only if

(31)V − 2κw

V≤(

2

d(L)

).


As d(L) gets large, the right-hand side converges from above to 0 and so in the limit, the

condition for stability becomes V ≤ 2κw, which is violated by our regularity condition. Thus,

there exists a finite d(L) such that the efficient network L is stable if and only if d(L) ≤ d(L).

Rearranging equation 31, L is stable if and only if

(32) d(L) ≤ 2

(V

V − 2κw

).

So the key threshold is d(κw) = b2V/(V − 2κw)c.Fixing the number of agents |N| in an efficient (tree) network L, the star network is the

unique (tree) network (up to a relabeling of players) that minimizes the diameter d(L) while

the line network is the unique (tree) network (up to a relabeling of players) that maximizes

the diameter d(L). The result now follows immediately.

Part (ii): On any efficient networks all links are essential and generate a net surplus

of V − 2κw > 0, where the inequality follows from our regularity condition. As i and j

must benefit equally at the margin from the link lij (see condition (ii) in the definition of

agreements that are robust to split the difference renegotiation), agent i’s expected payoff on

an efficient network L is

(33) ui(L) = |N(i;L)|(V/2− κw) > 0.

Thus i’s net payoff is proportional to his degree.

For any tree network L other than the star network let agent k be one of the agents with

the highest degree. Consider a link lij ∈ L such that i, j 6= k. As L is a tree there is a unique

minimal path from i to k and a unique minimal path length from j to k. As we are on a tree

network, either the path from j to k passes through i, or else the path from i to k passes

through j. Hence either i or j is closer to k and without loss of generality we let i have a

longer minimal path to k than j. We now delete the link lij and replace it with the link lik.

This operation generates a new tree network. Moreover, repeating this operations until there

are no links lij such that i, j 6= k, defines an algorithm.

This algorithm terminates at star networks as the operation cannot be applied to this

network; There are no links of lij such that i, j 6= k. Moreover the operation can be applied

to any other tree network because on all other tree networks there exists an lij such that

i, j 6= k. Finally, in each step of the algorithm the degree of k increases and so the algorithm

must terminate in a finite number of steps. Moreover, the algorithm must terminates at the

star network with k at the center.

By construction, at each step of the above algorithm we decrease the degree of some agent

j 6= k and increase the degree of k. Suppose we start with a network L and consider a step of

this rewiring where the link lij is deleted and replaced by the link lik. Only the expected payoff

of agents j and k on L and L∪lik \ lij change; The degrees of all other agents remain constant

and thus by equation 33 so do their payoffs. Letting α = (V/2−κw), we have uj(L) = αdj(L),

uk(L) = αdk(L), uj(L ∪ lik \ lij) = α(dj(L)− 1) and uk(L ∪ lik \ lij) = α(dk(L) + 1).


It follows that welfare W (u) =∑

i f(ui) (see equation 8) decreases through the rewiring

in this step if and only if

(34) f(α(dj − 1)) + f(α(dk + 1))− f(αdj)− f(αdk) < 0,

which is equivalent to:

(35) f(α(dk + 1))− f(αdk) < f(αdj)− f(α(dj − 1))

As f(·) is increasing, strictly concave and differentiable f ′(αdj)α < f(αdj)− f(α(dj − 1))

and f ′(αdk)α > f(α(dk + 1)) − f(αdk). Moreover, by concavity f ′(αdj) ≥ f ′(αdk) (as

dk ≥ dj). Combining these inequalities establishes the claim that f(α(dk + 1)) − f(αdk) <

f(αdj)− f(α(dj − 1)).

Thus at each step of the rewiring welfare W (u) decreases. For each network L′ reached

during the algorithm we can consider the average expected utility u′(L′) which if distributed

equally would generate the same level of welfare as obtained on L. As aggregate welfare is

decreasing at each step of the rewiring u′(L) must be decreasing too. However, the total

surplus generated by risk sharing remains constant and so average expected utility u remains

constant. Recall that Atkinson’s inequality measure / index is given by I(L) = (1−(u′(L)/u).

Thus at each step of the rewiring the inequality measure I(L) increases. As this rewiring

can be used to move from any tree network to the star network, stars network and only star

networks maximize inequality among the set of tree networks, which correspond to the set

of efficient networks under our regularity condition. As this argument holds for any strictly

increasing and differentiable, concave function f it holds for all inequality measures in the

Atkinson class.

Consider now an alternative rewiring of a tree network L. Let k be one of the agents with

highest degree on L and let j be one of the agents with degree 1 on L. As tree networks

contain no cycles, there always exists agents with degree 1 (leaf agents). Pick one of k’s

neighbors i ∈ N(k;L), remove the link lik from L and add the link lij to L. This operation

generates a new tree network. Repeating this operation until the highest degree agent has

degree 2 defines an algorithm. As the unique tree network with a highest degree of 2 is the

line network, the algorithm terminates at line networks and only line networks. At each stage

of the rewiring we either reduce the degree of the highest degree agent k or reduce the number

of agents who have the highest degree. Thus the algorithm must terminate in a finite number

of steps at a line network. Moreover, reversing the argument above, inequality is reduced at

each step of the rewiring for any inequality measure in the Atkinson class.

Proof of Proposition 7. By definition, underinvestment within group for a network L requires

that there exists an lij 6∈ L such thatG(i) = G(j) and for which TS(L∪lij)−TS(L) > 2κw. As


TS(L∪lij)−TS(L) = 0 for all non-essential links, lij must be essential on L∪lij. Thus lij is

also essential on L∪lij for any L ⊆ L. Equation 12 then implies that TS(L∪lij)−TS(L) ≥V for any L ⊆ L.

Consider any arrival order in which i arrives after j and let Si be the agents that arrive

(strictly) before i. Agent i’s marginal contribution to total surplus without lij when i arrives

is then TS(L(Si∪i))−TS(L(Si)) while with lij it is TS(L(Si∪i)∪lij)−TS(L(Si)). So

i’s additional marginal contribution to total surplus when lij has been formed is TS(L(Si ∪i) ∪ lij) − TS(L(Si ∪ i)). As L(Si ∪ i) ⊆ L, by the above argument TS(L(Si ∪i) ∪ lij) − TS(L(Si ∪ i)) ≥ V . As i arrives after j in half the arrival orders, i’s

average additional incremental contribution to total surplus when lij has been formed is

at least V/2. Thus MVi(L ∪ lij) −MVi(L) ≥ V/2. An equivalent argument establishes

that MVj(L ∪ lij) −MVj(L) ≥ V/2. Under our regularity condition V/2 > κw and so

i and j have a profitable deviation to form lij and the network L is not stable. As L was

an arbitrary network within underinvestment within group, there is no stable network with

underinvestment within group.

Proof of Proposition 8. The proof of the first part of the statement has four steps.

Step 1: Consider any efficient network L that is robust to overinvestment inefficiency

within group. This implies that for all path connected agents i, j such that G(i) = G(j)

and lij 6∈ L, either MVi(L ∪ lij) − MVi(L) ≤ κw or MVj(L ∪ lij) − MVj(L) ≤ κw.

However, by condition (i) in the definition of agreements that are robust to split the difference

renegotiation, MVi(L ∪ lij) −MVi(L) = MVj(L ∪ lij) −MVj(L) and so both MVi(L ∪lij)−MVi(L) ≤ κw and MVj(L ∪ lij)−MVj(L) ≤ κw.

Step 2: Let a network L := lij : G(i) = G(j), lij ∈ L be a network formed from L by

deleting all across-group links. Consider any subset of agents S ⊆ N such that i, j ∈ S. As

the network L is efficient, it is a tree network that minimizes the number of across-group

links conditional on a given set of agents being in a component. This implies that the unique

minimal path between i and j cannot contain an across-group link. So, i is path connected

to j on the induced subnetwork L(S) if and only if i is path connected to j on the induced

subnetwork L(S). Thus, by equation 12, the additional variance reduction that i and j

can now achieve by forming a superfluous across-group link on L(S) is weakly lower than

on L(S). So, by the Myerson value definition (equation 5), MVi(L ∪ lij) − MVi(L) ≤MVi(L ∪ lij) −MVi(L) and MVj(L ∪ lij) −MVj(L) ≤ MVi(L ∪ lij) −MVi(L). This

implies that L is robust to overinvestment within group.

Step 3: Let a network L′ be a network formed from L by rewiring (alternately deleting

then adding a link) each within-group network into a star (for an algorithm that does this,

see the part (ii) of the proof of Proposition 6). Consider any two agents i′, j′ such that

G(i′) = G(j′), li′j′ 6∈ L′. By part (i) of Proposition 6, MVi′(L′ ∪ li′j′) − MVi′(L

′) ≤


MVi(L ∪ lij) −MVi(L) and MVj′(L′ ∪ li′j′) −MVj′(L

′) ≤ MVj(L ∪ lij) −MVj(L).

Thus L′ is robust to overinvestment within group.

Step 4: Finally, consider any network L′ ∈ LCSS . This network can be formed by adding

a set of across-group links to a network L′ such that L′ ⊆ L′ and if lkk′ ∈ L′ \ L′ then

G(k) 6= G(k′). Consider any subset of agents S′ ⊆ N such that i′, j′ ∈ S′. Recall that

G(i′) = G(j′) and note that by the construction of L′, li′j′ 6∈ L′. On the induced subnetwork

L′(S′), either i′ is path connected to j′, in which case li′j′ would be superfluous if added,

or else i′ and j′ are isolated nodes. This is because the within-group network structure for

group G(i′) is a star. Thus, whenever li′j′ would not be superfluous, the change in i′ and

j′’s Myerson value if it were added is independent of the across-group links that are present:

MVi′(L′∪li′j′)−MVi′(L

′) = MVi′(L′∪li′j′)−MVi′(L

′) and MVj′(L′∪li′j′)−MVj′(L

′) ≤MVj′(L

′ ∪ li′j′)−MVj′(L′). Thus L′ is robust to overinvestment within group.

We turn now to the second part of the result. If L 6∈ LCCS , then there will be agents

i, j such that G(i) = G(j) and lij 6∈ L such that either the within-group network structure

for G(i) is not a star, or else it is a star but there are across-group links being held by an

agent who is not the center agent. In the first case, the inequality in step 3 will be strict

by Proposition 6. In the second case, we can without loss of generality let agent i be the

non-center agent holding the across-group link. Then, by equation 12, the inequality in step 2

will be strict. Thus for some parameter values L will not be robust to overinvestment within

group, but L′ will be.

Proof of Lemma 9. Denote the set of all possible arrival orders for the set of agents N, by

A(N). Order this set of |N|! arrival orders in any way, denoting the kth arrival order by

Ak ∈ A(N). We will then construct an alternative ordering, in which we denote the kth

arrival order by Ak ∈ A(N), such that for arrival order Ak,

(i) i arrives at the same time as agent i′ does for the arrival order Ak;

(ii) when i arrives he connects to exactly the same set of agents from N \ S0 that i′

connects to upon his arrival for the arrival order Ak;

(iii) when i arrives he connects to weakly more agents from S0 that i′ connects to upon

his arrival for the arrival order Ak.

Equation 14 shows that the risk reduction, and hence the marginal contribution made by

an agent k ∈ S0 from providing the across-group link lkj , is an increasing function of the

component size of k’s groups. It then follows that

(36) MV (i;L ∪ lij)−MV (i;L) > MV (i′;L ∪ li′j)−MV (i′;L).

To construct the alternative ordering of the set A(N) as claimed we will directly adjust

individual arrival orders, but in a way that preserves the set A(N). First, for each arrival

order, we switch the arrival positions of i′ and i. This alone is enough to ensure that conditions


(i) and (ii) are satisfied. There are |S0|! possible arrival orders for the set of agents S0.

Ignoring for now the other agents, we label these arrival orders lexicographically. First we

order them, in ascending order, by when i arrives. Next, we order them in ascending order

by the number of agents i is connected to upon his arrival. Breaking remaining ties in any

way, we have labels 1i, 2i, . . . , |S0|!i. We then let every element of A(N) inherit these labels,

so that two arrival orders receive the same label if and only if the agents S0 arrive in the

same order. We now construct a second set of labels by doing the same exercise for i′, and

denote these labels by 1i′ , 2i′ , . . . , |S0|!i′ . We are now ready to make our final adjustment

to the arrival orders. For each original arrival order Ak we find the associated (second)

label. Suppose this is xi′ . We then take the current kth arrival order (given the previous

adjustment), and reorder (only) the agents in S0, so that the newly constructed arrival order

now has (first) label xi. Because of the lexicographic construction of the labels, the arrival

position of agent i will not change as a result of this reordering of the arrival positions of

agents in S0, so conditions (i) and (ii) are still satisfied. In addition, condition (iii) will now

be satisfied from the definition of i being more central than i′. The only remaining thing

to verify is that the set of arrival orders we are considering has not changed (i.e. that we

have, as claimed, constructed an alternative ordering of the set A(N)) and this also holds by

construction.

Proof of Proposition 10. Let L be an efficient network that is robust to underinvestment

across group. This implies that for any across-group link lij ∈ L between groups g = G(i)

and g = G(j) 6= g, MVi(L) −MVi(L \ lij) = MVj(L) −MVj(L \ lij) ≥ κa, where the

inequality follows from condition (i) in the robustness to split the difference renegotiations

definition.

We now rewrite L. As the network L is efficient, it is a tree network that minimizes the

number of across-group links conditional on a given set of agents being in a component. This

implies that the unique minimal path between any two agents from the same group cannot

contain an across-group link. We can therefore rewrite the within-group network structures of

L to obtain a star by sequentially deleting and then adding within-group links (an algorithm

that does this is presented in the proof of part (ii) of Proposition 6). Do this rewiring so that

agent i is the agent at the center of the within-group network for group G(i) and let j be the

agent at the center of the within-group network for group G(j). Finally, we rewire across-

group links so that the same groups remain directly connected, but all across-group links are

held by the center agents. Let the network obtained be L′. By construction, L′ ∈ LCCS .

Under our definition of Myerson centrality, it is staightforward to verify that both i and

j are weakly more Myerson central within their respective groups on network L′ than on

network L. An argument almost identical to that in the proof of Lemma 9 then implies that

i′ and j′ have better incentives to keep the link li′j′ than i and j have to keep the link lij

(because the argument is more or less identical we skip it). Hence,


MVi′(L)−MVi′(L \ li′j′) ≥ MVi(L)−MVi(L \ lij)(37)

MVj′(L)−MVj′(L \ li′j′) ≥ MVj(L)−MVj(L \ lij)(38)

Network L′ is therefore robust to underinvestment. Moreover, whenever the within-group

networks of i and j on network L are not both stars with i and j at the centers, the inequality

is strict because both i and j are strictly more Myerson central within-group on L′ than on

L. There then exists a range of parameter specifications for which any center-connected star

network L′ ∈ LCCS is robust to underinvestment across group but L is not.

Proof of Proposition 11. We simply substitute si = α and si = β for i = 0, . . . , k into equation

12. This yields

∆ Var(L ∪ lij , L) =

(1− ρw) +2(k + 1)3β2α2 − 2(k + 1)3α2β2(∑ki=0 si

)(∑ki=0 si

)(∑ki=0 si + si

)(ρw − ρa)

σ2

= (1− ρw)σ2,

Multiplying by λ/2 to get the certainty-equivalent value of the variance reduction completes

the proof.

Proof of Proposition 13. We will say that agent k is a distance-t neighbor of i if the shortest

path from i to k take exactly t steps (and contain t+ 1 agents, including i and k).

Consider the implementation of the Incoming Link Deletion algorithm to find qij . We

begin by calculating x1,i = Aei, where ei is the ith basis vector. This identifies all agents

connected to i. We then set all entries in the ith row from the adjacency matrix A to 0 and

call this new matrix A2. This deletes the inward links toi in the network L. Starting from

i’s neighbors, we then find their neighbors on A2. In other words we calculate x2,i = A2x1,i.

This identifies the distance-2 neighbors of i. We then delete the rows of A2 that are indexed

by one of i’s neighbors, and so on.

In the tth round the algorithm identifies the distance-t neighbors of i. Thus, for t < l,

xl,ij = 0; for t = l, xl,ij = 1; and for all t > l, xt,ij = 0. Deleting incoming links ensures for all

t > l+ 1, xt,ij = 0. As L is a tree there, there is no path of length l+ 1 to j and so xl+1,ij = 0.

The algorithm therefore finds the unique minimal path from any i to any j and records its

length; If the unique path from i to j has length l, qij = 1/l. From equation 10 it is also

easily checked that qij = 1/l.


Appendix D. Income Correlation from Occupation Correlations

Here we present the rationale for our two caste–income correlation measures. Because we

lack income data, we must use occupation data in order to construct proxies. In our sur-

veys we have occupation data for all surveyed individuals, coded as small business owner,

land-owning farmer, farm laborer, dairy producer/cattle rearer, sericulture owner, sericul-

ture laborer, government official, garment worker, industrial factory worker, industrialist,

mason/construction worker, street vendor, artist (e.g., sculptor), and domestic help.

Let yi,g be the income and oi,g ∈ O the occupation of person i in group g ∈ A,B.Also, denote the probability that person i is in occupation o and j is in occupation o′ by

Pr(oi,A = o, oj,B = o′), where g(i) = A and g(j) = B. Finally, it will be useful to denote by

φg the proportion of individuals in group g.

In order to operationalize the quantity ρw − ρa in our empirical exercises, we use the

following measures:

(1) ρw − ρaI

=∑

g φgPr(oi,g = oj,g)− Pr(oi,A = oj,B)

(2) ρw − ρaII

=∑K

k=1 corr (Caste,Occupationk) Pr (Occupationk).

The first measure looks at the difference in the (weighted) share of pairs in the same caste

who hold the same occupation relative to the share of pairs in different castes who hold the

same occupation. We show that this measure is exact when (i) each individual draws an

income independently, although the mean can depend on caste and occupation, and (ii) there

are occupation level shock, and these shock has the same variance for every occupation.

The second measure presents a score that ranges from 0 to 1. If caste fully explains

occupation (where there is therefore scope for maximal within-caste income correlation), the

score is 1. However, if caste does not explain occupation at all, the score is 0. This measure

is computed as the average caste–occupation correlation, averaged over the occupations.

In what follows, assume the following:

(39) yi,g = µg,o + εo + ui,

where εo is a mean-0, variance-σ2o iid shock that hits each occupation, ui is an iid shock with

mean zero and variance σ2u that hits each individual, and µg,o is a (caste–occupation)-specific

mean.

By the law of total covariance, we have

cov(yi,A, yj,B) =∑o∈O

∑o′∈O

cov(yi,A, yj,B|oi,A, oj,B)Pr(oi,A = o, oj,B = o′)(40)

+cov(E[yi,A|oi,A],E[yj,B|oj,B]),

which we will use in our computations below.

D.1. Identical means and variances. We begin with the simple case where all means and

variances are identical. This justifies the use of ρw − ρaI.


Lemma D.1. Assume that µg = µ for all g and σ2o = σ2 for all o ∈ O. Then

(1) ρw ∝∑

g φgPr(oi,g = oj,g) and

(2) ρa ∝ Pr(oi,g = oj,g′),

both with the same constant of proportionality: σ2

σ2+σ2u

.

Proof. It is immediately clear that

cov(yi,A, yj,B|oi,A = o, oj,B = o′) = 0

if o 6= o′, and if o = o′ then

cov(yi,A, yj,B|oi,A = o, oj,B = o) = E[(εo)(εo)] = σ2.

Also, note that

cov(E[yi,A|oi,A],E[yj,B|oj,B]) = 0,

implying

cov(yi,A, yj,B) = Pr(oi,A = oj,B)σ2.

Thus the correlation between yi,g and yj,g′ is

corr(yi,g, yj,g′) =cov(yi,g, yj,g′)√

(σ2 + σ2u)(σ2 + σ2

u)

= Pr(oi,g = oj,g′)σ2

σ2 + σ2u

.

Weighting by population share, we have

ρw = φAρw,A + φBρw,B,

which completes the proof.

D.2. Differing average incomes by caste and occupation.

Lemma D.2. Assume that µg,o is allowed to vary by caste and occupation. Also, assume

that σ2o = σ2, for all o ∈ O. Then

(1) ρw ∝∑

g φgPr(oi,g = oj,g) and

(2) ρa ∝ Pr(oi,g = oj,g′),

both with the same constant of proportionality: σ2

σ2+σ2u

.

Proof. The same argument as in the preceding lemma gives us the result. One can check that

the heterogeneity in means does not affect the covariance terms.

D.3. Differing variances by occupation. Suppose now that

yi,g = µ+ εo + ui,

where εo is a mean-0, variance-σ2o iid shock that hits each occupation. It follows that

cov(yi,A, yj,B|oi,A = o, oj,B = o′) = 0


if o 6= o′, and if o = o′ then

cov(yi,A, yj,B|oi,A = o, oj,B = o′) = E[(εo)(εo)] = σ2o .

We still get cov(E[yi,A|oi,A],E[yj,B|oj,B]) = 0.

Therefore, we have

cov(yi,A, yj,B) =∑o∈O

Pr(oi,A = oj,B = o)σ2o .

Thus the correlation between yi,A and yj,B is

corr(yi,A, yj,B) =

∑o∈O Pr(oi,A = oj,B = o)σ2

o(∑o∈O Pr(o|A)σo

) (∑o′∈O Pr(o′|B)σo′

) .For people in different castes, one person from caste A and one from caste B, we then have

the following expression for across-caste income correlation:

ρa =

∑o∈O Pr(o|A)Pr(o|B)σ2



)For two people in caste A, we have the following expression for within-caste income corre-

lation:

ρw,A =

∑o∈O Pr(o|A)2σ2


)2 .In sum,

ρw = Pr(A)



)2 + Pr(B)

∑o∈O Pr(o|B)2σ2

o(∑o∈O Pr(o|B)σo

)2 .This implies that

ρw − ρa = Pr(A)



)2 + Pr(B)

∑o∈O Pr(o|B)2σ2

o(∑o∈O Pr(o|B)σo

)2−

∑o∈O Pr(o|A)Pr(o|B)σ2



) .Note that we do not know the values of σo for individual occupations, so we are unable

to compute this measure directly from the data as we did in previous subsections. Thus,

we take two approaches. First, we proceed as before, using Measure I, which ignores the

variance heterogeneity. This is an admittedly imperfect proxy for the desired measure here.

Second, we treat caste as a binary random variable and occupation as a multinomial, and

therefore take an occupation-share weighted correlation between caste and every occupation.

This is also an imperfect measure, but one that captures the intuition that if occupation can

be very strongly predicted by caste, then the caste–income correlation should be higher. For

this reason we use two proxies for our target quantity in the analysis.


Appendix E. Robustness and Supplementary Results

We now present results without differencing the social graph.

E.1. Myerson distance as a function of income variability and correlation. Figure

E.1 presents results in the raw data. As predicted, villages variable individual incomes are

associated with lower md(LF ), and villages with more within-caste income correlation are

associated with higher md(LF ).

(a) Measure: md(LF ) vs σ2 (b) Measure: md(LF ) vs ρw

Figure E.1. Myerson distance in standardized units against the variabilityof rainfall at the village level and the within-caste income correlation metric,both in standardized units.

Table E.1 demonstrates the robustness of this graphical evidence in regression of md(LF )

on σ, ρw, and with district or subdistrict fixed effects as well as controls for caste composition.

Columns 1-3 present regressions using ρwI , whereas columns 4–6 present regressions using

ρw − ρaII

. Panel A presents results using the Incoming Link Deletion algorithm whereas

Panel B looks at Outgoing Link Deletion. We find a one standard deviation increase in

Measure I is associated with a 0.292 or 0.296 standard deviation increase in md(LF ) (column

1, Panels A and B). However, there is no significant relationship between income variability

and md(LF ) in these specifications. Column 4 presents the analogous results using ρw − ρaII

.

A one-standard-deviation in Measure II is associated with a 0.205 or 0.199 standard deviation

increase in md(LF ), but income variability has no detectable correlation with the outcome

variable (column 4, Panels A and B). In columns 2 and 5 we add district fixed effects, and the

results are mostly reflective of those in columns 1 and 3. A one standard deviation increase

in income correlation corresponds to a roughly 0.1 or 0.18 standard deviation increase in

Myerson distance, using Measure I (p-values 0.477 and 0.12 in Panels A and B) and Measure

II (p-values 0.00 and 0.012 in Panels A and B) respectively. Inclusion of district fixed effects

resolves some of the noise around the rainfall variability coefficient as well in the case of

Panel A. Finally, when we include subdistrict fixed effects, our estimate under Measure I is

too noisy to distinguish from zero, though we find similar evidence under Measure II (a 0.1

standard deviation effect, p-value 0.04 and 0.13 in Panels A and B). Further, when comparing

villages within subdistricts we are able to identify an income variability effect. A one standard


deviation increase in income-variability is associated with a 0.3 standard deviation decrease

in the average Myerson distance (columns 3 and 6, respectively, of both panels).

Taken together, we show an increase in within-caste income correlation is associated with

a change in Myerson distance in a manner consistent with our theory, though the evidence

for income variability is considerably weaker.

Table E.1. Myerson distance vs. income variability and correlation

(1) (2) (3) (4) (5) (6)

Income Correlation 0.292 0.086 -0.006 0.199 0.180 0.150(0.137) (0.106) (0.108) (0.092) (0.055) (0.079)[0.046] [0.477] [0.935] [0.034] [0.000] [0.044]

Income Variability 0.009 -0.217 -0.278 0.015 -0.278 -0.315(0.126) (0.131) (0.097) (0.158) (0.131) (0.094)[0.96] [0.154] [0.002] [0.951] [0.072] [0.002]

District FE N Y Y N Y YSubdistrict FE N N Y N N YR-squared 0.1335 0.4368 0.6728 0.0891 0.4637 0.6924

Income Correlation 0.296 0.109 -0.001 0.205 0.182 0.098(0.109) (0.061) (0.064) (0.088) (0.069) (0.063)[0.004] [0.12] [0.996] [0.03] [0.012] [0.13]

Income Variability -0.061 0.015 -0.281 -0.054 -0.054 -0.304(0.187) (0.239) (0.094) (0.206) (0.224) 0.101[0.82] [0.928] [0.002] [0.774] [0.794] [0.002]

District FE N Y Y N Y YSubdistrict FE N N Y N N YR-squared 0.1066 0.5701 0.7458 0.071 0.5897 0.7525

Measure I Measure II

Notes: Outcome variable is the average Myerson distance, computed by the approximation algorithms as indicated.Panel A uses the incoming link deletion algorithm and Panel B uses the outgoing link deletion algorithm. Outcomevariables and regressors scaled by their standard deviations. Columns (1-3) use Measure I of the caste-incomecorrelation, whereas columns (4-6) use Measure II. Specifications include control for caste composition: p(1-p), where p is share in high caste. Wild clustered bootstrap standard errors are presented in (.), using Rademacher weights. Thecluster is at the subdistrict level, of which there are 12. 1000 samples are used per bootstrap. p-values from the Wildclustered bootstrap t are presented in [.].

Panel A: Incoming Link Deletion Algorithm

Panel B: Outgoing Link Deletion Algorithm

E.2. Association between within-caste centrality and cross-caste links. Table E.2

presents the results using just the financial networks on regressors of interest. The main

result is that an increase in the measure of the ρw − ρa, irrespective of which measure is

used, corresponds to about a roughly 0.15 standard deviation decline in the average Myerson

centrality (within caste) of the individuals with across caste links. Again, when we look at

rainfall variability, we are unable to confirm or reject our theory, as estimates are very noisy.


Thus, our evidence is partially in support of our theory. We find a negative association

between income correlation and the centrality of those with a cross-caste link. However, we

are unable to confirm or reject our hypothesis when looking at the variability of income.

Table E.2. Association between average within-caste centrality of nodeswith cross-caste links and measures of rainfall variability and within-casteincome correlation


Income Correlation -0.231 -0.196 -0.131 -0.101(0.098) (0.142) (0.061) (0.097)[0.018] [0.157] [0.032] [0.28]

Income Variability 0.009 -0.019 0.086 0.093(0.183) (0.157) (0.185) (0.191)[0.961] [0.918] [0.636] [0.609]

R-squared 0.2265 0.1992 0.127 0.111

Incoming Link Deletion Algorithm Outgoing Link Deletion Algorithm

Notes: Outcome variable is the average within-caste average Myerson centrality, computed using our approximation algorithm (incoming linkdeletion or outgoing link deletion), of those with a cross-caste link. Income correlation given by Measure I or Measure II as indicated.Outcome variables and regressors are scaled by their standard deviations. The outcome variable uses only the financial network. Wild clusteredbootstrap standard errors are presented in (.), using Rademacher weights. The cluster is at the subdistrict level, of which there are 12. 1000samples are used per bootstrap. p-values from the Wild clustered bootstrap t are presented in [.].

social investments, informal risk sharing, and …

Documents